Journal Pre-proof

Recent Advancements in Machine Learning and Deep Learning for

Early Detection of Breast Cancer: A Comprehensive Review

Tangin Amir Smrity , MD ZAHIN MUNTAQIM ,

Hasan Muhammad Kaf

PII: S2950-2128(26)00001-1
DOI: [Link]
Reference: IBREH 100050

To appear in: Innovative Practice in Breast Health

Received date: 22 September 2025

Revised date: 7 December 2025
Accepted date: 23 January 2026

Please cite this article as: Tangin Amir Smrity , MD ZAHIN MUNTAQIM , Hasan Muhammad Kaf ,
Recent Advancements in Machine Learning and Deep Learning for Early Detection of Breast Cancer:
A Comprehensive Review, Innovative Practice in Breast Health (2026), doi: [Link]
ibreh.2026.100050

This is a PDF of an article that has undergone enhancements after acceptance, such as the ad-
dition of a cover page and metadata, and formatting for readability. This version will undergo addi-
tional copyediting, typesetting and review before it is published in its final form. As such, this ver-
sion is no longer the Accepted Manuscript, but it is not yet the definitive Version of Record; we are
providing this early version to give early visibility of the article. Please note that Elsevier’s sharing
policy for the Published Journal Article applies to this version, see: [Link]
policies-and-standards/sharing#4-published-journal-article. Please also note that, during the produc-
tion process, errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.

© 2026 The Author(s). Published by Elsevier Masson SAS.

This is an open access article under the CC BY-NC-ND license
([Link]
 1

Highlights

 Comprehensive review of Machine Learning (ML) and Deep Learning (DL) approaches for breast cancer
detection and classification (2020–2025).

 The study categorizes approaches into two major groups: ML-based and DL-based methodologies.

 ML-based approaches include traditional algorithms such as SVM, Random Forest, k-NN, and ensemble
techniques with feature engineering.

 DL-based approaches are further classified by imaging modality: ultrasound, histopathological images, thermal
imaging, and mammograms.

 Identification of current trends, key challenges (e.g., data imbalance, model interpretability), and limitations in
existing research.

 Discussion on potential future directions including explainable AI, data fusion, and real-time clinical integration.
 2

Recent Advancements in Machine Learning and Deep Learning for Early Detection of Breast Cancer: A
Comprehensive Review

Tangin Amir Smrity , MD ZAHIN MUNTAQIM , Hasan Muhammad Kaf

Corresponding Author: MD ZAHIN MUNTAQIM MUNTAQIM Bangladesh Army University of Science and
Technology BANGLADESH
 3

Abstract—Breast cancer remains one of the most prevalent and life-threatening diseases affecting women
globally. Early and accurate detection is crucial for effective treatment and improved survival rates. In recent years,
Machine Learning (ML) and Deep Learning (DL) techniques have shown significant promise in enhancing the
accuracy, speed, and reliability of breast cancer diagnosis and classification. This review presents a comprehensive
analysis and quality assessment of research studies published between 2020 and 2025, evaluating dataset
representativeness, reference standards, validation methodology, and risk of bias using adapted QUADAS-AI
Framework, focusing on ML and DL-based approaches applied to breast cancer detection and classi-fication. We
categorize the reviewed literature based on the type of input data (e.g., mammograms, histopathological images,
ultrasound, and clinical data), learning models (such as Support Vector Machines, Random Forests, Convolutional
Neural Networks, and Transformers), and performance metrics used. Additionally, we highlight key trends,
challenges, and innovations, including the rise of hybrid models, transfer learning, and explainable AI in medical
diagnostics. This review aims to provide researchers and clinicians with a structured understanding of the recent
advancements and future directions in AI-driven breast cancer diagnostics.

Index Terms—Breast Cancer Detection using Deep Learning, Breast Cancer Detection using Machine Learning,
Breast Cancer Detection using Computer Vision, Breast Cancer Detection Review, Ultrasound imaging,
Histopathological imaging, Mammography, AI in Healthcare

I. INTRODUCTION

Breast cancer continues to be a major public health concern and remains one of the leading causes of cancer-related
mortality among women globally [1]–[3]. Early detection and accurate diagnosis are crucial to improving patient
outcomes; however, traditional diagnostic methods largely reliant on radiologist expertise and manual interpre-tation
are inherently subjective and can lead to inconsistent results [4], [5]. The growing complexity and volume of medical
imaging data further underscore the need for automated, reliable, and efficient diagnostic tools.
In recent years, Artificial Intelligence (AI), particularly Machine Learning (ML) and Deep Learning (DL), has
emerged as a trans-formative force in medical diagnostics [6], [7]. These techniques have demonstrated exceptional
capabilities in learning intricate pat-terns from high-dimensional data sources such as mammograms, histopathological
slides, ultrasound scans, and electronic health records. The application of ML and DL not only enhances diagnostic
accuracy but also aids in reducing false positives and negatives, thereby supporting faster and more consistent clinical
decision-making [8], [9].
Between 2020 and 2025, there has been a surge in research focused on leveraging AI for breast cancer detection and
classification. Recent studies have proposed innovative architectures, including convolutional neural networks
(CNNs), ensemble models, and hybrid approaches that integrate clinical and imaging data for improved diagnostic
outcomes [7], [9]. These advancements reflect a clear shift towards data-driven methodologies that surpass traditional
rule-based systems in performance, scalability, and adaptability.
This review synthesizes key research contributions from 2020 to 2025, offering a comprehensive analysis of ML
and DL techniques
4

applied to breast cancer diagnostics. To provide a structured overview, we have divided the paper into two main
categories: Machine Learning (ML)-based approaches and Deep Learning (DL)-based approaches. In the ML
section, we explore conventional algorithms such as Support Vector Machines (SVM), Random Forests, and k-
Nearest Neighbors (k-NN), focusing on their applications, feature extraction techniques, and performance metrics. In
the DL section, we further categorize the studies based on the type of imaging modality used namely, ultrasound,
histopathological images, ther-mal imaging, and mammograms, highlighting how different neural network
architectures are optimized for each modality. Additionally, we discuss the datasets employed, evaluation metrics
reported, and the comparative effectiveness of these techniques. By examining current trends and identifying existing
limitations, this work aims to guide future research efforts toward more robust, interpretable, and clinically viable AI
solutions for breast cancer care.

II. METHODOLOGY

This study employed a systematic literature review approach to investigate the application of deep learning
techniques in breast cancer detection and classification. The review was structured to provide a thorough and critical
evaluation of the current body of research, emphasizing recent advancements, persisting challenges, and prospective
directions for future work in this domain.

A. Databases Searched

In order to comprehensively review the literature on AI models for breast cancer detection, we systematically
searched multiple electronic databases. These included:
• PubMed: A comprehensive database of biomedical literature, particularly focused on life sciences and clinical
studies.
• IEEE Xplore: A leading digital library for engineering and technology-related research.
• Scopus: A multidisciplinary database covering scientific, tech-nical, and medical fields.
• Google Scholar: For broader coverage, including grey literature and conference proceedings.
These databases were selected based on their relevance to the field of medical imaging, machine learning, and deep
learning applications.

B. Keywords and Search Strategy

We used the following keywords and phrases to guide our search:

• “Breast cancer detection using AI”
• “Machine learning for breast cancer diagnosis”
• “Deep learning in breast cancer imaging”
• “Breast cancer classification using neural networks”
• “AI models for mammography, ultrasound, thermography”
• “AI model validation in medical imaging”
• “Dataset issues in breast cancer AI”
 5

These keywords were combined with Boolean operators (AND, OR) to ensure a comprehensive search that
captured both specific

and broad aspects of AI in breast cancer detection. We also reviewed references from identified articles to ensure no
relevant studies were missed.

C. Inclusion and Exclusion Criteria

We established clear criteria to ensure that only the most relevant and high-quality studies were included in this
review:

Inclusion Criteria

• Studies published between 2020 and 2025, focusing on AI models (machine learning or deep learning) for breast
cancer detection using imaging modalities (e.g., mammograms, ultra-sound, thermography).
• Studies that include performance evaluation metrics (e.g., accu-racy, sensitivity, specificity, AUC).
• Peer-reviewed journal articles, conference papers, and clinical trial reports.
• Studies that provide sufficient methodological details for repro-ducibility and transparency.

Exclusion Criteria

• Studies that do not focus on AI-based models or breast cancer detection.

• Studies using non-medical imaging modalities or those unrelated to breast cancer.
• Non-peer-reviewed articles or abstracts without sufficient data or results.
• Studies that do not report on model validation, external valida-tion, or clinical relevance.
Figure 1 summarizes the number of papers after each step in paper searching and selection.

D. Risk of Bias Assessment

While evaluating the studies included in this review, we also considered the risk of bias across several domains as
outlined by QUADAS-AI, which is particularly pertinent to AI models in medical imaging. The following factors were
assessed:
• Dataset Representativeness: Many studies included in the review utilized datasets from specific hospitals or
institutions, which may not represent the broader patient population. The risk of bias increases when datasets are
not diverse in terms of patient demographics, imaging equipment, and clinical settings, limiting the
generalizability of the findings.
• Reference Standards: The quality and consistency of refer-ence standards (e.g., pathologist annotations, biopsy
results, or mammography classifications) are critical in determining the accuracy of AI models. Studies that did not
use well-established, standardized reference benchmarks may introduce significant bias in model performance
6

evaluation.
• Data Leakage: In some studies, there is a potential risk of data leakage during model training, where information
from the test data may influence the training process. This can lead to inflated performance metrics that do not
reflect the model’s true clinical applicability.

E. Synthesis of Evidence Quality

The studies reviewed were classified into three categories based on their quality: high, moderate, and low, each with
varying degrees of reliability and generalizability.
 7

Fig. 1: PRISMA Flow Diagram of Our Study Selection Process.

• High Quality Studies: These studies generally used diverse datasets, often involving multi-center data, which
enhance gen-eralizability. For example, the study by Jungkyu et al. (2025)
[10] utilized a large, multi-center dataset of over 750,000 exams, ensuring robust results that are likely to
generalize to broader populations. Additionally, high-quality studies performed thor-ough external validation,
achieved consistent results across mul-tiple datasets, and employed explainable AI (XAI) techniques to improve
clinical interpretability.
• Moderate Quality Studies: These studies often faced limita-tions like small sample sizes, often from single
centers, which may affect the generalizability of the results. For example, Proshenjit et al. (2025) [11] used a
small clinical dataset (116 samples), which can result in overfitting and reduce the model’s applicability to larger
populations. Many of these studies also lacked external validation, which is essential for confirming the clinical
8

relevance and robustness of the models.

• Low Quality Studies: A significant portion of studies fell into this category due to issues like lack of external
validation, overfitting to small or curated datasets (e.g., Wisconsin dataset),and biased reporting due to artificial
class balancing. Studies those used the BreakHis dataset faced issues of patient-level overfitting, where models
performed well only on specific cases but failed to generalize. These issues lead to an overestimation of the
models’ performance in clinical settings.

III. M ACHINE LEARNING-BASED APPROACHES

Machine learning (ML) has emerged as a central tool in breast can-cer research, with studies exploring a wide
spectrum of algorithms, data modalities, and interpretability strategies. A consistent theme across recent works is the
dual pursuit of predictive accuracy and clinical relevance.
Chowdhury et al. [12] introduced an interpretable ML framework for classifying breast cancer subtypes using RNA-
sequencing data. They applied dimensionality reduction and optimized multiple models via grid search, ultimately
improving classification performance. No-tably, they integrated SHAP values to identify transcriptomic features that
were both predictive and biologically meaningful, adding a level of interpretability often missing in prior models.
Tested on TCGA data, their approach outperformed existing methods across several metrics, including accuracy,
precision, and F1-score. Furthermore, gene set enrichment analysis linked top features to known molecular pathways,
reinforcing the biological and clinical relevance of the findings.
Building on the need for both performance and explainability, another study [11] proposed a feature selection-based
classification method that enhanced prediction accuracy with a reduced feature set. Using a wrapper-based approach
combined with three metaheuristic algorithms: Whale Optimization, Bald Eagle Search, and Sea Lion Optimization
paired with an XGBoost classifier, the authors applied the method to the Breast Cancer Coimbra dataset. Each
algorithm reduced the original nine features to four, while maintaining or improving performance. The best-performing
model, SLOA XGB, achieved an F1-score of 97.43%, with 97.14% accuracy and 100% re-call. SHAP analysis
identified glucose as the most influential feature, and correlations between certain feature means and misclassification
rates provided additional interpretive insights.
Focusing on imaging data, Li et al. [13] developed binary logistic regression models to predict each breast cancer
molecular subtype such as Luminal A, Luminal B, HER2, and TNBC, using multimodal ultrasound features combined
with clinical data. Models incorporating all ultrasound modalities (conventional, shear wave elastography, and contrast-
enhanced) outperformed those based on single modalities. The HER2 subtype model achieved the highest accuracy,
with an AUC of 0.89.
Mono-omics data also served as a foundation in the study by Hassan et al. [14], which evaluated several classifiers:
random forest, partial least squares, Naive Bayes, decision trees, neural networks, and Lasso regression on data from
the Cancer Genome Atlas. Among these, random forest and Lasso achieved the highest AUC scores (0.99), and
consistently outperformed others across multiple metrics. The study highlighted the strength of these models for early
breast cancer diagnosis based on genomic profiles.
Addressing recurrence prediction, Zuo et al. [15] compared eleven ML algorithms and identified AdaBoost as the
top-performing model using metrics such as AUC, accuracy, sensitivity, and F1-score. SHAP values were employed to
 9

interpret feature importance, highlighting CA125, CEA, fibrinogen (Fbg), and tumor diameter as key predictors. This
was the first study to apply SHAP for enhancing interpretability in breast cancer recurrence prediction with AdaBoost,
offering a practical decision support tool for clinicians.
Gradient boosting models were evaluated in a related effort by Nahid et al. [16], who compared Gradient Boosting
(GB), XGBoost,
10

CatBoost, and LightGBM for classifying patients using clinical features such as glucose and insulin levels. The GB
model achieved the highest accuracy (82.85%) and F1-score (84.21%). SHAP values, combined with filter and
wrapper-based selection methods and ag-gregated via Borda count, revealed glucose as the most influential feature.
Misclassified samples were further examined to identify sources of model error.
Incorporating imaging data and patient records, Senthil et al. [17] evaluated models including CNN, KNN, RNN,
SVM, and random forest. The CNN model performed best, achieving 98.7% accuracy, with strong results across
other metrics as well. This underscored the potential of deep learning in image-based breast cancer diagnostics and
the benefits of integrating diverse model architectures.
A broader comparative study [18] evaluated eight classifiers on the clinical dataset, consistently ranking SVM
highest in accuracy (97%+), followed by MLP and stacking methods. RF also performed well but trailed SVM.
Similarly, Karatza et al. [19] explored in-terpretability techniques like SHAP and ICE plots in ensembles, showing
improved performance after feature selection, with RF’s accuracy rising from 96.49% to 97.18%.
The use of machine learning (ML) and Explainable AI (XAI) has significantly advanced breast cancer
diagnosis and prediction by improving both accuracy and interpretability. Building on this, one prominent study [20]
investigates the use of ML and XAI to improve breast cancer classification. The study employed five supervised ML
models such as decision tree, random forest, logistic regression, naive Bayes, and XGBoost on a real-world clinical
dataset of 500 patients from Dhaka Medical College Hospital. The evaluation met-rics included classification
accuracy, precision, recall, and F1 scores. XGBoost emerged as the best performer, achieving an impressive
accuracy of 97%. Additionally, SHAP (SHapley Additive exPlana-tions) analysis was applied to interpret the
predictions of the XGBoost model, offering insights into the influence of individual features on the model’s decision-
making process. This approach demonstrated that XGBoost not only improved classification performance but also
enhanced the trustworthiness of the breast cancer diagnosis model.
Another study by Arravalli et al. [21] focused on identifying breast cancer using machine learning classifiers and
incorporating XAI for model transparency. The study evaluated several ML algorithms, with Random Forest
achieving the best performance, recording an F1-score of 84%. The study also explored the use of a stacked ensemble
model, which combined multiple classifiers to improve prediction accuracy. The stacked model yielded an F1-score
of 83%. To improve model interpretability, several explainers, including SHAP, LIME, ELI5, Anchor, and QLattice,
were applied, helping clinicians understand the factors influencing the model’s predictions. This combination of ML
and XAI enhanced model transparency, making it a valuable tool for clinical decision-making in breast cancer
diagnosis.
A third study presented a tailored approach for Indian breast types [22], focusing on features such as breast
density, texture, lesion size, and composition, which differ from international datasets. Machine learning algorithms
were optimized for mammographic images from Indian populations. The best-performing individual model achieved
an Area Under the Curve (AUC) of 0.95, and by combining four models, the AUC improved to 0.98. Results from an
independent test set demonstrated 86.7% sensitivity and 96.1% specificity, proving the effectiveness of the deep
learning-based approach in enhancing diagnostic accuracy for the Indian population. This study highlights the
importance of considering population-specific characteristics to improve breast cancer detection [22].
Further investigation into the histological subtypes of Invasive Breast Cancer (IBC) focused on differentiating
 11

between Invasive Ductal Carcinoma (IDC) and Invasive Lobular Carcinoma (ILC) by analyzing glandular texture
features from pre-contrast T1-weighted MRI images. [23] The clinical dataset was processed using 3D Slicer for
segmentation. Texture features were extracted using MATLAB, including first-order statistics and GLCM (Gray
Level Co-occurrence Matrix) features. Feature selection via the ANOVA F-test identified correlation (0.1233) and
mean (0.5335) as the least significant features. Despite this, an initial model with all features showed high
accuracy (0.9038), emphasizing the importance of retaining all extracted features. To address class imbalance, the
SMOTE technique was applied, resulting in a balanced dataset with 80% training and 20% testing. The Random
Forest classifier achieved the highest cross-validation scores, with an accuracy of 91% on the original dataset and 87%
on the SMOTE dataset. This study emphasizes the use of MRI texture features and the importance of class balance in
improving the accuracy of early breast cancer subtype diagnosis.
Tegaw et al. [24] investigated the performance of machine learning models for predicting breast cancer survival,
specifically based on treatment modalities like chemotherapy, hormone therapy, surgery, and radiation therapy. A
dataset consisting of 50000 samples from Kaggle was used to evaluate multiple classifiers, including Support Vector
Machines (SVM), K-Nearest Neighbor (KNN), AdaBoost, Gradient Boosting, Random Forest, and Extreme Gradient
Boosting (XGBoost). Gradient Boosting emerged as the best performer, with an accuracy of 97.2%, precision of 0.973,
recall of 0.972, F1-score of 0.973, and an AUC of 0.997. SHAP analysis was used to explain feature impacts on
predictions, revealing that chemotherapy had a negative effect on survival, while hormone therapy showed the most
detrimental impact. Radiation therapy had the most positive impact on survival outcomes. This study highlights the
importance of Gradient Boosting in predicting survival and provides actionable insights into the effects of different
treatments on patient survival.
A different approach combined traditional parametric survival models with machine learning algorithms to predict
breast cancer survival outcomes. The study by Kaindal et al. [25] applied log-gaussian regression and logistic
regression models alongside neural networks, SVM, Random Forest, and Gradient Boosting Machines (GBMs) to
assess the relationship between survival and clinical variables. The neural network model achieved the highest
predictive accuracy, while the Random Forest model demonstrated the best balance between fit and complexity, as
indicated by the lowest Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) values. The
study found that age, tumor grade, AJCC stage, marital status, and radiation therapy use were significant predictors of
survival. This study emphasizes the value of integrating traditional survival analysis techniques with machine learning
approaches for improved predictive accuracy.
The potential of ML was further explored in predicting survival outcomes for Metaplastic Breast Cancer (MBC), a
rare and aggressive breast cancer subtype [26]. Using data from the SEER database (2010-2018), the researchers
applied CatBoost to predict 1-year, 3-year, and 5-year survival outcomes, achieving AUC values of 0.833, 0.806, and
0.810, respectively. The model maintained strong performance when validated on an independent external dataset. The
study also assessed the impact of radiotherapy, finding it to be most beneficial for certain patient groups, such as those
undergoing breast-conserving surgery with M0 stage or mastectomy with T3-4/N2-3M0 stage. The study
concluded that CatBoost is a valuable tool for predicting survival outcomes in MBC patients and for optimizing
treatment strategies based on cancer stage.
A key challenge addressed in breast cancer research is predicting the conversion of HER2-0 primary breast cancer to
12

a HER2-low phenotype at relapse. This is crucial for determining eligibility for anti-HER2 antibody–drug
conjugates in metastatic breast cancer treatment. In the study [27], Two ML models were trained: one explainable
(XGBoost) and the other performance-driven (ensemble of XGBoost and SVM). The explainable model achieved a
balanced accuracy of 58%, with sensitivity of 53% and specificity of 64%. The ensemble model improved
performance, achieving a balanced accuracy of 64%, with higher sensitivity (75%) but lower specificity (53%). This
study represents one of the first applications of ML to predict HER2-low conversion, offering new possibilities for
treatment access and the integration of AI in breast cancer oncology.
Koume et al. [28] focused on Fear of Cancer Recurrence (FCR) in breast cancer survivors developed an ML model
to predict patients at risk of developing clinical FCR five years after their diagnosis. The dataset from the VICAN
survey included 918 patients, and the study employed several ML algorithms, including Random Forest, SVM,
Gradient Boosting, eXtreme Gradient Boosting, and Multilayer Perceptron. The best-performing model achieved an
AUC of 66%, showing the potential of ML algorithms in predicting FCR. While the results were promising, the
study emphasized the need for further refinement to enhance model accuracy and generalizability, ultimately
contributing to personalized prevention strategies for survivors.
Table I provides a comprehensive summary of research studies employing machine learning approaches, outlining
their key contri-butions, reported results, identified limitations, and clinical relevance.

IV. DEEP LEARNING-BASED APPROACHES

A. Thermographic Data
Thermography has gained significant attention as a non-invasive, radiation-free alternative for breast cancer
detection, particularly advantageous for patients with dense breast tissue or in settings where mammography is
inaccessible [32]. Over the years, research has evolved from traditional texture-based classifiers to sophisticated deep
learning systems, focusing on improving accuracy, interpretability, and practical deployment.
Early efforts in this domain established the feasibility of thermal imaging for breast cancer detection. For example,
Acharya et al. [33] extracted texture features from thermal images using co-occurrence and run-length matrices and
applied a Support Vector Machine (SVM) classifier, achieving 88.10% accuracy on a small dataset. Although limited
in scale, this study demonstrated the potential of texture-based ML for thermal image classification.
Building upon these foundations, recent works have introduced advanced deep learning architectures and hybrid
models. Dharani et al. [34] proposed the Enhanced Deep learning-based Convolu-tional Neural Network (EDCNN),
which used heatmaps to assess vascularity and detect hotspots in thermal breast images. The model achieved 96.8%
accuracy and 93.7% specificity on the Graphical DMR dataset, outperforming conventional methods while offering
greater interpretability. Similarly, Bani Ahmad et al. [35] designed a custom deep learning framework,
StackVRDNet, which integrated features from VGG16, ResNet, and DenseNet. Their use of the novel Rock
Hyraxes Dandelion Algorithm Optimization (RHDAO) for segmentation and parameter tuning led to a 97.05%
accuracy, showcasing the potential of ensemble and optimization strategies.
Other studies have prioritized interpretability alongside accuracy. Aidossov et al. [36] addressed the “black-box”
issue in deep learning by combining CNN outputs with Bayesian Network (BN) reasoning. While MobileNet alone
achieved 93.8% accuracy, the hybrid CNN-BN model offered better interpretability with minimal performance loss
(90.85% accuracy). Similarly, Alshehri et al. [37] improved CNN performance by integrating attention mechanisms
 13

(AMs), which allowed the network to focus on thermally significant regions. Their CNN-AM model outperformed
traditional CNNs with test accuracies reaching up to 99.46%.
 14

TABLE I: Summary of Breast Cancer Machine Learning Studies: Contributions, Results, and Quality Assessment.

Study Contribution Result Study Limitations Dataset Reference Dataset Issues Clinical Risk of Bias
Quality Representativ Standards Relevance
e-
ness
Chowdhur et al Interpretable Outperforme High Limited Limited to Clear Dataset Highly relevant Low risk of bias;
y . ML d existing generalizabilit TCGA, may (TCGA limited to for under- used
(2025) [12] framework for meth- y across not represent dataset) TCGA, standing tumor SHAP for feature
breast cancer ods on different diverse pa- may not subtypes and interpre-tation
subtype TCGA data; populations tient groups represent guiding
classification improved broader personalized
using accuracy, demographics treatments
RNA-seq; precision,
SHAP values F1-score;
for feature identified
interpretability biologically
and gene set mean-ingful
enrichment features
analysis
Proshenjit et al. Feature Reduced Moderat Limited feature Small sample Clear Small dataset; Medium relevanc Moderate risk of
[11] selection features e set size in the (clinical limited fea- for e bias due
combining from 9 to clinical dataset dataset) ture set, could clinical practice to dataset size
wrapper 4; best affect gen- especiall ,
methods model eralization y in feature
with SLOA XGB reductio decisio
metaheuristic achieved n for n
algorithms and 97.43% F- support
XGBoost score, systems
classifier 97.14%
accuracy,
100%
recall;
Glucose
identified as
most
influential
feature
 15
Li et al. (2025) Multimodal HER2 High Potential Small dataset Clear Small sample High relevance Low risk of bias;
[13] ultrasound subtype variability in from a sin- (multimod size and for used
and clinical model imaging gle hospital al data limited multimodal multimodal data
data logistic achieved techniques sources) diversity of diagnostic sources
regression highest ultra-sound approaches in
models for AUC of techniques clinical settings
breast cancer 0.89; mul-
subtype timodal
prediction models
outperforme
d single-
modality
models
Hassan et al. Evaluation of RF and High Mono-omics Potential Clear Dataset lacks High relevance for Low risk of bias;
(2025) multiple ML Lasso focus biases in (TCGA variety; lim- person- validatio robust
[14] models on achieved may limit TCGA dataset) its alized medicine, n technique
mono-omics highest biological dataset generalizabilit particu-larly in s
TCGA data; AUC 0.99; interpretation y to di-verse predicting breast
random for-est consistently populations cancer subtypes
and Lasso best
regression precision,
identified as recall, and
best F1
Zuo et al. Compared 11 AdaBoost High Small sample Relied on Clear Limited Medium Moderate risk of
(2023) [15] ML algo- outperforme size may limited clinical (clinical dataset, could clinical bias due
rithms for d oth- affect markers markers) skew clinical relevance; needs to sample size
breast cancer ers; SHAP generalizability (CA125, CEA) relevance broader
recurrence identified validation in
prediction; first CA125, diverse cohorts
use of SHAP CEA,
with Ad-aBoost fibrinogen,
tumor diam-
eter as top
predictors
Nahid et al. Gradient GB model Moderat Possible Limited Clear Small sample Medium Moderate risk of
(2022) boosting mod- best: 82.85% e overfitting sample size, (clinical size; dataset clinical bias due
[16] els with SHAP accu- due to not dataset) not relevance; needs to model
and com-bined racy, hyperparamete representative representative clinical trial complexity
feature 84.21% F1; r tuning of larger of di-verse validation to
selection via Glucose top populations clinical confirm findings
Borda count predictor; settings
method analysis of
 16
misclas-
sified
samples
Senthil et al. Compared CNN highest High Limited Strong Clear Imbalance in High clinical Low risk of bias;
[17] CNN, KNN, accuracy clinical multimodal (CT/MRI dataset can relevance, strong
RNN, SVM, 98.7%; validation; data data) lead to skewed especially for focus on technical
RF on strong focused on usage, but results real-world evalu-ation
CT/MRI and precision, technical dataset imbal- multimodal
medical records recall, F1; performance ance diagnostic
for breast highlights systems
cancer multimodal
prediction data ad-
vantage
Aroef et al. Applied Boruta SVM High Potential bias Small dataset Clear Dataset size High relevance High risk of bias
(2024) feature se- accuracy in fea- (small and selection for im- due to
[29] lection before 95%, RF ture selection dataset) process may proving small dataset
RF and SVM 90%; process affect robust- classification ac-
training feature ness curacy in clinical
selection diagnos-tics
significantly
impacted
performance
Taminul et al. ML and XAI for XGBoost High Limited Dataset Clear Dataset not High relevance Low risk of bias;
(2025) improv- achieved generalizabilit limited to (Real-time representative for im- XAI
[20] ing accuracy highest ac- y across Dhaka hospital of broader proving techniques used
and inter- curacy of populations Medical dataset) demographics diagnostic trans-
pretability of 97%; SHAP College parency
breast can-cer used for Hospital
classification; interpretabilit
com-pared y
decision tree,
ran-dom forest,
logistic re-
gression, naive
Bayes, and
XGBoost
Arravalli et al. ML classifiers Random High Overfitting risk Limited Clear Small dataset, High clinical Low risk of bias;
(2025) for breast Forest F1- with dataset; may (UCTH might affect relevance for XAI for
[21] cancer score 84%; ensemble not clinical generalization decision-making interpretability
prediction with ensemble model represent dataset)
XAI for model F1- larger popula-
interpretability; score 83% tions
 17
evaluated
Random Forest,
stacking
ensemble
Puttegowd et al Optimized ML AUC of 0.98 High Limited to Limited to Clear Limited by High relevance Low risk of bias;
a . algorithms with four- Indian Indian popu- (DDSM regional focus for im- diverse
(2025) [22] for Indian model breast types; lation; may not INbreast proving models used
breast types integration; may not represent clini- diagnostic accu-
using high generalize global cal dataset) racy in specific
mammographic sensitivity globally diversity popula-tions
im-ages and
specificity
on
independent
dataset
Nuzla et al. ML model for Achieved High Class Dataset Clear Imbalance in High relevance Low risk of bias;
(2025) differenti- 91% imbalance imbalance; (Cancer dataset could for early SMOTE
[23] ating IDC and accuracy on addressed may Imaging affect accuracy diagnosis of breast used for class
ILC using MRI original with affect model Archive; cancer subtypes balancing
texture features dataset; SMOTE; performance Frederick
Random feature National
For-est selection Laborator
classifier needed y for
performed improvement Cancer
best Research
(FNLCR))
LaMoglia e al ML classifiers Logistic High Overfitting Small dataset; Clear Limited High relevance Moderate risk of
(2025) [30] t . to improve Regression risk in not diverse (clinical generalizabilit for im- bias due
breast cancer achieved non-feature enough dataset) y proving breast to feature selection
diagnosis; 91.67% selected cancer di-agnostic
compared with accuracy; models methods
traditional feature se-
methods like lection
mammogra-phy improved
LGBM per-
formance to
90.74%
Kaindal et al. ML for breast Neural High High Complex Clear High model High relevance Low risk of bias;
(2025) cancer network complexity of model integra- (SEER complexity for sur- validatio robust
[25] survival model integrating tion may clinical may affect vival prediction n technique
prediction, achieved traditional and affect inter- data) clinical adop- and per-sonalized s
combining highest ML models pretability tion treatment
 18
traditional predictive
models with accuracy;
machine Random
learning Forest
model best
balance of
fit
Tegaw et al. ML models for Gradient High Focused on Dataset Clear Focus on High relevance Low risk of bias;
(2025) predicting Boosting treatment limited to (clinical treatment, lim- for sur- SHAP
[24] breast cancer achieved modalities, not treat- dataset) iting vival prediction used for
survival out- highest dis-ease ment variables generalizabilit and treat-ment interpretability
comes based on performance progression y for broader optimization
treatment ; SHAP survival
modalities analysis predic-tions
revealed
insights into
treatment
impacts
Zhang et al. CatBoost for AUC of High Limited Dataset Clear Limited to High relevance Low risk of bias;
(2025) predicting 0.833 for 1- to specific to (SEER MBC, may not for rare data SEER
[26] survival year Metaplastic MBC; database) apply to other subtype survival used
outcomes in survival, Breast might not be breast can-cer predic-tion
Metaplastic strong Cancer; may applicable to types
Breast performance not generalize all breast
Cancer; across to other types cancer types
assessment of datasets
radiotherapy
benefits
Miglietta et al. ML to predict Ensemble High Focused on Dataset limited Clear Focused on High relevance Low risk of bias;
(2025) HER2-0 XGBoost HER2- to HER2- (Data from specific con- for robust
[27] to HER2-low and SVM low 0 patients; 5 Italian version, limits optimizing HER2- methodology
conversion at achieved conversion, might not in- generaliz- targeted therapies
relapse; focus 64% may not apply to all stitutions) ability
on anti-HER2 balanced generalize to cases
treatment accu-racy; other BC types
eligibility demonstrate
d AI poten-
tial for
improving
treatment
access
Koume et al. ML for Best model High Limited Healthcare Clear Reimbursemen Moderate Moderate risk of
 19
(2025) predicting Fear achieved accuracy, reimbursement (VICAN t data relevance for bias;
[28] of Cancer 66% further data; may not ((VIe might miss post-cancer care dataset limitations
Recurrence in AUC; refinement capture full apre`s le psychological
breast cancer demonstrate needed patient CANcer)) factors
survivors d potential experience dataset) affecting FCR
for FCR
prediction
Priyadarsh et al ML for breast Combined High Limited to Dataset limited Clear Limited by High relevance Low risk of bias;
ni . cancer deep CBIS- to screen- (CBIS- dataset focus for im- robust
(2024) [31] detection using learning and DDSM; may ing DDSM) on proving model integration
CBIS-DDSM traditional not generalize mammography mammograph diagnostic accu-
dataset; ML to other images y images racy
combined deep techniques datasets
learning and achieved
traditional ML superior
techniques accuracy
20

In terms of practical deployment and real-time analysis, several studies proposed mobile-friendly and efficient
pipelines. AlHusaini et al. [38] developed a thermal video-based detection system using mobile-connected FLIR
cameras. Their modified Inception Mv4 model achieved an impressive 99.748% accuracy. They also found that the
application of cooling gel significantly enhanced thermal contrast and diagnostic sensitivity. Similarly, Jalloul et al.
[39] emphasized both accuracy and efficiency by combining ResNet152 with an SVM classifier, achieving 97.62%
accuracy with minimal computational latency (0.06 s), thus validating the feasibility of real-time thermographic
analysis.

Several works also focused on classification granularity, moving beyond binary classification. Attallah et al. [40]
introduced Thermo-CAD, which achieved perfect accuracy (100%) in normal vs. abnor-mal classification and 79.3% in
benign vs. malignant classification. Their pipeline used multiple CNNs for feature extraction, followed by
dimensionality reduction with non-negative matrix factorization and feature selection via Relief-F. Khan et al.
[41] conducted a rare direct comparison between thermography and mammography, demonstrating that thermography
achieved comparable performance in normal vs. abnormal classification (96.57% vs. 98.11%) and superior results in
benign vs. malignant classification (92.70% vs. 82.05%). These findings underscored thermography’s potential in
resource-constrained environments.

Collectively, these studies demonstrate a clear evolution in ther-mographic breast cancer diagnosis, from early
SVM-based models to complex, interpretable deep learning frameworks. Innovations in feature extraction,
attention mechanisms, and hybrid reasoning have significantly enhanced model performance. At the same time,
increasing emphasis on explainability and real-time deployment has brought thermal imaging closer to practical,
clinical utility. As models continue to improve in accuracy, interpretability, and accessibility, thermography stands
out as a powerful complementary tool in breast cancer screening.

Table II provides a comprehensive summary of research studies employing deep learning approaches, outlining their
key contribu-tions, reported results, identified limitations, and clinical relevance.

B. Mammogram Data

Mammography remains one of the most widely used modalities for early breast cancer screening and diagnosis.
Recent advances in artificial intelligence have significantly improved the accuracy, efficiency, and interpretability of
mammogram analysis. A range of approaches have emerged, from single-view lesion segmentation and molecular
subtype prediction to large-scale multi-modal systems and hybrid deep learning pipelines.
One of the most impactful developments is the shift from analyzing single mammographic views to leveraging
paired or multi-view data. Seo et al. [50] introduced PMVnet, a paired-view U-Net-based architecture that models
the relationships between standard mammographic views (e.g., CC and MLO). By incorporating co-sine similarity
and a squeeze-and-excitation mechanism, the model achieved a Dice Similarity Coefficient of 0.709 and a recall of
0.950, substantially outperforming single-view baselines. This inter-view relational modeling marked a step forward in
 21

reducing false positives and improving segmentation quality.

Alongside structural improvements, researchers have explored the use of attention mechanisms to boost diagnostic
accuracy and inter-pretability. Zeng et al. [51] enhanced ResNet18 with a Convolutional Block Attention Module
(CBAM) to predict HER2, ER, and PR status directly from mammograms. Without requiring tumor segmentation, the
model achieved promising AUCs, particularly for ER (0.785), and delivered strong F1 scores, showing potential for
noninvasive subtype classification. Similarly, Isarun et al. [47] employed a modi-fied EfficientNetV2 with attention
mechanisms to detect malignancy in mammograms. Their model achieved AUROCs up to 0.96 and demonstrated high
acceptance rates among expert radiologists, rein-forcing AI’s potential as a supportive tool in clinical workflows.
Hybrid data and feature fusion approaches have also gained trac-tion. Murty et al. [52] combined tabular data from
the clinical Breast Cancer Database with mammographic imaging data (CBIS-DDSM), enabling the model to learn
both cellular and visual characteristics. Vijayalakshmi et al. [62] proposed another hybrid framework that merged
handcrafted statistical features (e.g., GLCM, GLRLM) with deep features in a BiLSTM-CNN model, achieving
97.14%. These approaches highlight the benefits of integrating multi-domain infor-mation for comprehensive breast
cancer characterization.
At the population scale, Jungkyu et al. [10] developed a multi-modal AI system incorporating full-field digital
mammography (FFDM), synthetic mammography, and digital breast tomosynthesis (DBT). Trained on nearly 500,000
exams and later scaled to 750,000, the system achieved an AUROC of 0.945 while reducing false-positive recalls by
31.7% and radiologist workload by 43.8%, without compromising sensitivity. This large-scale integration shows how
AI can enhance both diagnostic performance and clinical efficiency in screening programs.
Curriculum learning and supervision-level-aware training strategies have been proposed to address annotation
bottlenecks in real-world mammography datasets. Ozden et al. [48] introduced a patch-based learning model that
progressively increased supervision levels. With only partial bounding box annotations, the model still achieved high
F1 scores and improved explainability via Grad-CAM visualizations, demonstrating the viability of training
explainable models under sparse supervision.
For segmentation-focused tasks, AbdelWahab et al. [49] combined EfficientNet for multi-scale feature extraction
with U-Net for pixel-
22

level segmentation. Optimization through Particle Swarm Optimiza-tion (PSO) enabled the framework to reach over
98% classification accuracy and 91% IoU and Dice scores on CBIS-DDSM dataset. The method outperformed
several standard baselines while remaining computationally efficient.
Finally, ensemble learning has been explored to further boost performance. Chakravarthy et al. [63] proposed a
fuzzy ensemble of three deep CNNs: VGG-11, Inception v3, and ResNet50 enhanced with a modified Gompertz
function for adaptive score integration. This approach achieved a classification accuracy of 98.99%, outper-forming
traditional ensemble techniques like weighted averaging.
Taken together, these contributions reflect a rich and evolving land-scape in AI-driven mammography analysis.
From inter-view model-ing and attention mechanisms to large-scale multi-modal integration and hybrid feature
fusion, researchers are pushing the boundaries of accuracy, interpretability, and scalability. These innovations not
only support radiologists in early cancer detection but also pave the way for more personalized and efficient
screening paradigms.

C. Histopathological Data

Histopathological analysis remains the gold standard for definitive breast cancer diagnosis. With the rise of
computational pathology, a wide range of deep learning-based models have emerged, aiming to improve accuracy,
reduce pathologist workload, and provide timely and reproducible diagnostic support. These methods span from con-
ventional CNN-based classification to attention-enhanced architec-tures, hybrid models, biomarker-level analysis,
and explainable AI systems.
A major focus in recent work has been on designing architectures that balance diagnostic accuracy with
computational efficiency. For instance, Alzakari et al. [64] proposed the Deep Sparse Wavelet Autoencoder
(DSWAE), which integrates wavelet-based feature ex-traction and sparse coding within an autoencoder framework.
By reducing the number of trainable parameters while preserving critical morphological features, DSWAE achieved
high classification preci-sion, particularly 100% for normal tissue, minimizing false positives and making it well-
suited for real-time deployment.
Building on deep convolutional approaches, several studies ex-plored the use of state-of-the-art backbone
architectures for feature extraction. A DenseNet121-based model [53] demonstrated strong performance in
distinguishing benign and malignant breast tissue from biopsy slides, while Naren et al. [65] achieved peak classifi-
cation accuracy of 99.01% on high-magnification histopathological images using an EfficientNetV2-XL architecture
enhanced with a Convolutional Block Attention Module (CBAM). These improve-ments were attributed to both
attention mechanisms and preprocessing techniques such as CLAHE, which sharpened the visibility of relevant tissue
features.
Attention-based hybrid models have also gained traction. Pradeepa et al. [60] proposed a deep learning framework
combining Effi-cientNetV2 for feature extraction with a GRU enhanced by at-tention mechanisms to model
spatial and sequential dependencies. Their approach achieved strong results across multiple metrics on Camelyon17
datasets. Similarly, Singh Patnaik et al. [66] introduced Patho-AI, a robust ensemble-based system incorporating
models like ResNet50 and ConvNeXtTiny, and augmented with Explainable AI techniques such as Grad-CAM,
 23

LIME, and Integrated Gradients. This combination not only enhanced diagnostic performance but also addressed a
critical clinical requirement: interpretability.
In parallel, researchers have explored segmentation-classification pipelines to improve feature localization and
lesion-specific modeling. Potti et al. [67] introduced a two-stage pipeline that uses Fuzzy
 24

TABLE II: Summary of Deep Learning-based Breast Cancer Studies.

Category Stud Contribution Results Study Limitations Dataset Reference Dataset Clinical Risk of Bias
y Qualit Representative Standards Issues Relevance
y ness

[42] Hybrid CNN + Vision F1-score: High None reported Moderate; Clear Potential Highly Moderate risk

Ultrasound Trans- 95.57%, Accu- small, single- (BUSI bias from relevant for of bias due
former model for racy: 95.63%, site dataset) small, ultrasound- to small
breast ultra-sound Sensitiv- dataset single-site based breast dataset
image analysis ity: 96.42%, dataset cancer
(BUSI). Precision: detection
94.79%

[43] Developed 2 novel InBnFUS Modera Limited Limited; small Clear Small Medium High risk of
architec- gains 99% te sample size sample size (clinical dataset; relevance for bias due to
tures: InBnFUS and accu- dataset) limited breast dataset size
CNNDen-GRU. racy. generalizabi ultrasound
lity analysis

[44] Developed a AUC: 0.968, High Potential Moderate; Clear Small High Low risk of
multimodal Accuracy: variability in multimodal (mammograp sample size, relevance bias; used
framework that 93.78%, imaging integra- hy and limited for multimodal
integrates features Specificit techniques tion ultrasound) diversity in multimodal data sources
from mammography y: datasets diagnostic
and ultrasound. 96.41%, approaches
Precision
:
83.66%
 25

[45] Evaluation of Best accuracy: High None reported High; strong Clear Small High Low risk of
pretrained ResNet50 feature (BUSI dataset; relevance bias; good
CNNs (ResNet50, at 98.41% extraction dataset) may not for representative
MobileNet, VGG16) from ultrasound generalize ultrasound- ness
for breast ultrasound to broader based breast
classification. clin-ical cancer
settings detection

[46] Predict BI-RADS AUROC 0.854 Modera Potential Moderate; Clear Potential Medium Moderate risk
density from for den- te variability in potential (clinical for relevance for of bias due
handheld BUS sity; 5-year handheld variability dataset) variability cancer risk to small
images; AI-derived risk prediction ultrasou in handheld in handheld prediction dataset
density as a cancer AUROC nd imaging image quality ultrasound
risk factor. 0.633 imaging

[47] Modified AUROCs: High Single Moderate; Clear Limited to Medium Low risk of
EfficientNetV2 with 0.89, 0.96, dataset; single dataset (clinical one dataset; relevance for bias; expert
attention for 0.94; Expert reduces dataset) re- mammogram validation
Mammogra mammogram can-cer acceptance generalizabilit duces analysis
m detection and lesion >89% y generalizabi
local-ization. lity

[10] Multi-modal AI AUROC High Dataset High; multi- Clear Dataset High Low risk of
system 0.945; 31.7% imbalance; modal data inte- (FFDM, imbalance; relevance bias; strong
combining FFDM, re- poten- gration synthetic poten- for validation
synthetic duction in tial biases in mammograp tial biases multi-modal
mammography, and false image types hy, DBT) in image screening
DBT for screening. positives; types approaches
43.8%
workload
reduc-
26
tion; 100%
sensitivity

[48] Patch-based F1 improved Modera Patch-based Moderate; Clear Patch-based Medium Moderate risk
mammogram de- from 80.55% te data; limited patch-based (mammo dataset; relevance for of bias due
tection with to 83.95% with dataset data gram may im- to dataset
curriculum learn-ing, full super- diversity dataset) not proved quality
weak/strong vision represent mammogram
supervision, Grad- full clinical anal-ysis
CAM explainability. data
diversity

[49] EfficientNet + U-Net Accuracy High None reported High; well- Clear Dataset size High Low risk of
for tumor 98.7% (CBIS- balanced dataset (CBIS- limits relevance for bias; robust
segmentation/classific DDSM) IoU DDSM gener- auto- methodology
ation; 91%, Dice Dataset) alizability mated tumor
hyperparameters >91% detection
optimized by Particle
Swarm.

[50] Introduces PMVnet, a Dice Modera None reported Moderate; Clear Small High Low risk of
paired- Similarity te single-view and (mammo dataset; relevance bias; novel
view mammogram Coeffi- paired-view gram may not for model design
network designed to cient of 0.709, integration dataset) generalize mammogram
improve breast le-sion recall: 0.95 to diverse -based
detection. clin-ical detection
settings

[51] Integration of AUC: 0.708 Modera None reported Moderate; Clear Limited High Low risk of
Convolutional for HER2, te focus on (mammo focus on relevance bias; attention
Block Attention 0.785 for ER HER2, gram spe- for mechanism
Module to ResNet18. and 0.706 for ER, PR dataset) cific marker-based applied
 27
PR. classification markers; detection
generaliza-
tion may be
affected

[52] Integration of CNNs The model Modera Dataset Moderate; Clear ( Imbalanced Medium Low risk of
with fine- learns high- te imbalance; dataset CBIS- dataset relevance for bias
tuning on the CBIS- level visual may imbalance DDSM) affects mammogram
DDSM dataset. patterns. lead to generalizati diagnosis
overfitting on

[53] DenseNet121 CNN Superior High None reported High; large Clear May not High Low risk of
for breast accuracy, dataset from di- (histopatholo represent all relevance bias; diverse
cancer subtype preci- verse sources gy sub- for dataset

Histopathol classification on sion, recall, dataset) types histopatholog

ogy histopathology and F1-score y analysis

images. vs. existing
methods.

[54] Automated TILs Cohen’s kappa High None reported High; strong Clear (TILs Potential High Low risk of
assessment in 0.71 agree- correlation with dataset) biases in relevance for bias; high
QuPath combining ment with pathologist TIL auto- agreement
pixel clas-sification pathologists. results classification mated TIL
and cell detection. assessment

[55] Ensemble framework, 90.1% High No High Clear Moderate High Moderate risk
aggre- accuarcy implementatio (Clinic Class relevance of bias; lack
gating SVM, RF, n of Ex- al Imbalance for of
XGBoost, DT, and plainable AI CIAR2018 e ensemble- explainability
AdaBoost (xAI) BACH based
Challeng diagnostic
Dataset) methods
28

[56] HI-Net 90.4% High No High Clear Lack of High Moderate risk
accuarcy implementatio (Combin Diversity relevan of bias
n of Ex- ation ce
plainable AI of for
(xAI) CAMELYO histopath
N 16, ology
CAMELYO classification
N 17 &
ICIAR 2018
Clinical
Datasets)

[57] multimodal deep AUC of 0.72 High None reported High Clear Lack of High Low risk of
learning in TCGA (TGCA Generalizati relevan bias; excellent
method data, 0.76 in Clinical on to ce reproducibility
two-fold Dataset + Non-Cancer for
cross- Cancer Datasets histopath
validation Hospital of ology
the Chinese classification
Academy of
Medical
Sciences)

[58] Breast U-shaped 91.2% average High None reported High Clear Lack of High Low risk of
Network accuracy, (TNBC diversity in relevance for bias; excellent
(BU-NET) framework 91.8% average Clinical TNBC TNBC reproducibility
F1 Dataset) dataset diagnosis

[59] EfficientNetB3 98.39% High None High Clear May not High Low risk of
Accuracy Reported (combi represent relevan bias; excellent
ned low- ce reproducibility
 29
LC25000 magnificati for
and GDC on data histopath
Cancer Data ology
Portal classification
dataset)

[60] Hybrid Accuracy High None reported High; large Clear Imbalanced High Low risk of
EfficientNetV2 + 95.72%, multi-class (Clinic data may relevance for bias; high pre-
attention-based GRU Precision dataset al af- multi- cision
for spatial/sequential 98.15%, Recall Camelyon1 fect class
modeling in 95.68%, F1- 7 detection classification
histopathology. score 96.82% histopatholo
gy dataset)

[61] Deep learning for Best accuracy Modera Focused on Moderate; focus Clear Focus on High Low risk of
Ki-67 98.8%, in- te specific on biomarker (biopsy specific relevance bias; specific
biomarker ference time biomarker images) biomarker for detection
segmentation 13 seconds. limi biomarker
and classification in ts detection
biopsy images. generalizati
on

[34] EDCNN framework Accuracy High None reported Moderate; Clear Potential Medium Low risk of

Thermograp for ther- 96.8%, Speci- single-site data (thermal IR dataset relevance for bias; novel

hic mal IR breast images; ficity 93.7%; images) bias; thermal IR approach
fuzzy clustering + outperforms variation breast
heatmap vascular-ity existing across imaging
analysis. thermal patients
methods.

[33] Support vector Accuracy: Modera Small dataset Clear (thermal Small dataset Medium Moderate risk Moderate risk
machine classi- 88.10%, Sen- te dataset) may reduce relevance of bias due of bias due
30
fier. sitivity: generalizabili for to small to small
85.71%, ty thermal IR dataset dataset
Speci- detection
ficity: 90.48%

[40] Introduced Thermo- Accuracy: Modera None reported Moderate; Clear Small Medium Low risk of
CAD that 79.3% te thermal imaging (thermal dataset, relevance for bias; good
integrates multiple images) limited thermal IR validation
CNNs. clinical detection
validation

[41] Proposed a Gabor Accuracy: High None reported Moderate; Clear Single-site High Moderate risk
feature and 96.57% single dataset (thermal data; limits relevance for of bias;
SVM-based dataset) clinical CAD dataset
computer-aided applicability systems in limitations
detection (CAD) diagnosis
system.

[36] A comparative study MobileNet High None reported High; diverse Clear High risk of High Low risk of
of pre- Accuracy: models (thermal bias due to relevance for bias; good
trained state-of-the-art 93.8%, BN: dataset) model ther- performance
DL models, 90.85% selection mography
Developed Bayesian Accuracy diagnostics
Network.
 31

C-Means clustering for segmentation followed by a U-Net-based CNN for classification. This setup significantly
outperformed SVMs, achieving 99.23% accuracy. Likewise, Arun Kumar et al. [68] pro-posed a transfer learning-
based framework combined with traditional classifiers such as SVM and KNN, using PCA to boost both accuracy and
runtime efficiency, achieving 99.5% accuracy while reducing execution time, highlighting its practicality for resource-
constrained clinical settings.

Another novel direction involves biomarker-level histopathology analysis. Thoung et al. [61] developed a deep
learning pipeline targeting the Ki-67 protein index, a key biomarker of tumor cell proliferation. Leveraging models like
DeepLabv3+, DenseNet-121, and U-Net, the system accurately segmented and quantified Ki-67-positive tumor
cells and tumor-infiltrating lymphocytes (TILs), achieving diagnostic accuracy as high as 98.8% while maintaining low
inference time (13 seconds). This type of targeted biomarker analysis provides more granular insight into tumor
behavior and opens pathways for integrating AI with precision oncology.

Supporting such biomarker analysis, the work by Bankhead et al. [54] built a fully automated TIL quantification
pipeline within the open-source QuPath software. Using StarDist for cell detection and classical binary classifiers for
cell classification, the pipeline aligned well with pathologist annotations (Cohen’s kappa = 0.71), showing that open-
source, interpretable, and accessible tools can play a vital role in augmenting diagnostic workflows.

Finally, multi-class classification has also seen progress. Mudassir et al. [69] introduced PRMS-Net, which
combines Progressive Resid-ual Networks and ResNet-50, achieving 99.63% accuracy with robust performance
across precision, recall, and F1-score. This method showed a notable reduction in both false positives and false
negatives, an essential requirement for trustworthy diagnostic AI in clinical settings.

In summary, histopathological image analysis for breast cancer has rapidly evolved beyond simple classification.
Modern AI approaches integrate attention, explainability, biomarker-level segmentation, and hybrid architectures to
achieve not just high accuracy but also in-terpretability, efficiency, and clinical relevance. These advancements
collectively push the field toward scalable, trustworthy AI tools ca-
32

Fig. 2: PMVnet Architecture proposed by [50].

 33

pable of supporting and augmenting routine pathological assessments in diverse healthcare settings.

D. Ultrasound Data

Ultrasound imaging is widely recognized for its safety, acces-sibility, and affordability, particularly in breast
cancer screening and diagnosis. However, the modality’s inherent limitations such as operator dependency, variable
image quality, and ambiguous tumor boundaries have motivated the development of AI-driven solutions to improve
diagnostic consistency and accuracy. Recent deep learning approaches have addressed challenges in both classification
and segmentation, with innovations ranging from multimodal fusion to transformer-based architectures.
One notable direction has been the development of robust clas-sification models for breast ultrasound images
(BUSI). Fatima et al. [43] proposed two novel architectures, InBnFUS and CNNDen-GRU, designed for adaptability
across multiple ultrasound domains, including both breast cancer and maternal-fetal imaging. These archi-tectures fused
Inception modules with inverted bottlenecks and GRUs to capture both spatial and sequential features. The models
achieved 99.0% accuracy on breast cancer data, confirming their effectiveness in generalizing across heterogeneous
ultrasound datasets.
Complementing these developments, Aamir et al. [42] introduced CB-Res-RBCMT, a hybrid framework
combining CNNs and Vi-sion Transformers (ViTs) to capture both local textures and global morphological
variations in BUS images. The model integrated spatial attention and channel-boosting techniques to enrich feature
representation, achieving a high F1-score of 95.57%, outperform-ing conventional CNN and ViT baselines.
Similarly, Mumtahina et al. [45] evaluated standard CNN architectures (ResNet50, VGG16, and MobileNet) on a
large dataset, with ResNet50 leading in perfor-mance at 98.41% accuracy, reinforcing its role as a robust backbone
for ultrasound-based classification.
Segmentation, particularly of irregularly shaped tumors, remains a critical task for early and precise diagnosis.
Sharma et al. [70] addressed this by developing a multiscale feature fusion framework combined with transfer learning
and ensemble segmentation. Their model, tested across three public datasets, achieved high Dice scores (92.11% on
UDAIT), demonstrating strong generalization and the ability to accurately delineate even small lesions, an essential
step toward improved clinical utility in early-stage cancer detection.
Another emerging focus is the use of multimodal learning, combin-ing ultrasound with other imaging modalities.
Chen et al. [44] con-structed a deep learning framework that integrates mammography and ultrasound features,
demonstrating that multimodal models signifi-cantly outperformed single-modality systems. Their best-performing
model achieved an AUC of 0.968 and 93.78% accuracy, highlighting how cross-modal feature fusion can enhance
diagnostic precision. Notably, the model’s predictions were interpretable through heatmap visualizations, validating its
clinical relevance.
Beyond traditional classification and segmentation, ultrasound is also being explored for breast density estimation
and risk stratifi-cation. Arianna et al. [46] trained deep learning models on over 400,000 BUS images to predict BI-
RADS breast density categories. The best model achieved an AUROC of 0.854, and AI-derived ultrasound density was
shown to predict 5-year breast cancer risk with nearly the same accuracy (AUROC = 0.633) as mammographic BI-
RADS density. This suggests that deep learning models can extract meaningful radiomic information from ultrasound,
34

expanding its role beyond lesion detection to broader risk modeling.

In summary, ultrasound-based breast cancer analysis is rapidly evolving with the integration of advanced deep
learning methods. Recent models are addressing the challenges of image heterogeneity, weak contrast, and limited data
through novel architectural designs, attention mechanisms, multimodal fusion, and transfer learning. As these models
continue to improve in accuracy, generalizability, and interpretability, they show strong potential for real-world clinical
adoption, offering low-cost, non-invasive, and scalable solutions for early breast cancer detection and monitoring.
Table II provides a comprehensive summary of research studies employing deep learning approaches, outlining their
key contribu-tions, reported results, identified limitations, and clinical relevance. The table III provides a comparative
analysis of the different modal-ities.

V. ADDRESSING LIMITATIONS

A. Dataset Limitations

While datasets such as WBCD [19], [71]–[78], MIAS [62], and BreakHis [65]–[69] have been foundational in early
research, it is important to note that they do not fully represent the complexity and variety found in clinical settings.
These datasets often suffer from limitations such as small sample sizes, single-site collection, and class imbalances,
which can lead to overly optimistic performance metrics that do not generalize to broader, real-world applications.

B. Key Clinical Problems Unsolved by AI

Despite the significant advancements in AI-driven breast cancer detection, several key clinical challenges remain
unresolved. These challenges include:
1) Early Detection of Subtle Cancers:
AI systems have demonstrated strong performance in detecting more obvious or well-defined cancers, but early-
stage cancers, particularly those that manifest as subtle abnormalities or have ambiguous features, remain
difficult to detect [79]. This limitation results in a higher risk of false negatives, which can delay diagnosis
and treatment, ultimately affecting patient outcomes [80].
2) Integration with Clinical Workflow:
The adoption of AI in clinical practice continues to face significant hurdles, primarily due to the lack of
seamless integration with existing radiological workflows. AI tools often provide diagnostic support, but they
fail to work collaboratively within the decision-making process of clinicians, leading to inconsistent adoption
and effectiveness in real-world clinical settings [81].
3) Handling Data Variability:
AI models trained on specific datasets from single institutions often struggle to generalize across diverse clinical
environ-ments, where imaging protocols, equipment types, and patient demographics can vary significantly
[82]. This data variability results in models that may perform well on curated datasets but fail to deliver
consistent results in real-world scenarios.
4) Interpretability and Trust:
One of the primary barriers to AI adoption in clinical breast cancer detection is the lack of transparency in the
 35

decision-making process of AI models. The “black-box” nature of deep learning algorithms makes it
difficult for clinicians to understand why certain predictions are made [83], [84], which undermines the trust
needed for widespread clinical adoption. Clinicians require interpretable AI systems that can explain their
reasoning and align with medical knowledge to facilitate integration into clinical workflows.

VI. CHALLENGES

A. Clinical Workflow Integration in Screening vs. Diagnostic Settings

One of the key challenges in the clinical adoption of AI models is their integration into existing workflows,
particularly given the differences between screening and diagnostic settings. In screening environments, AI models
must prioritize sensitivity to ensure that as many potential cases as possible are flagged for further examination,
minimizing the risk of missed diagnoses. However, this approach can lead to an increased number of false
positives, as the models tend to err on the side of caution. In contrast, diagnostic settings require AI systems to
prioritize specificity to reduce unnecessary procedures such as biopsies or additional imaging, which can lead to
patient anxiety and healthcare resource strain. Therefore, the integration of AI into clinical workflows should be
tailored to these distinct goals: in screening, AI models must help radiologists identify potential cases early, while in
diagnostic settings, the focus should shift toward confirming or ruling out malignancy with high confidence. The
different diagnostic needs of each setting influence the design of AI systems and their evaluation metrics. For
example, in screening applications, a higher sensitivity is preferable, even at the cost of reduced specificity, while in
diagnostic settings, the trade-off should favor reducing false positives and ensuring that the identified lesions are
truly malignant.

B. BI-RADS and AI in Clinical Breast Cancer Detection

The Breast Imaging Reporting and Data System (BI-RADS) is a key tool used by radiologists to standardize
the interpretation of mammographic findings and to guide clinical management. AI models, particularly those using
machine learning and deep learning, have the potential to augment or even enhance BI-RADS assessments by
providing additional quantitative data to support qualitative human evaluations. AI systems can assist in identifying
subtle patterns that
36

Fig. 3: pipeline of multimodal model [44].

TABLE III: Comparative Analysis of Imaging Modalities for Breast Cancer Detection
Dimension Mammography Ultrasound Histopathology Thermography Synthesis
Clinical Role
Avg. Reported Accuracy
Primary screening Diagnostic adjunct for dense breasts Gold standard diagnosis Emerging screening alternative Gap: No modality-specific AI validation protocols exist
External Validation Rate
Computational Cost
Explainability Implementation 96.8% (range: 89–99%) [10], [47]–[52], [62] 95.9% (range: 93.78–99%) [42]–[46] 98.2% (range: 91–100%) [53], [54], [60], [61], [65]–[69] 94.3% (range: 88–97%) [33], [34], [36], [40], [41] Finding: Thermography underperforms despite being radiation-free

14% 0% 8% 0% Major gap: 95% of studies lack multi-center validation

High (ResNet50: 25M params) Moderate (EfficientNet: 10M params) Very High (ensemble models: 50M+ params)
29% 12.5% 23% Low (MobileNet: 3M params) Practical barrier: Real-time deployment requires model compression

12.5% Urgent need: 82% of studies lack clinically-aligned explainability

may be missed by human observers, thereby improving the accuracy and consistency of BI-RADS scoring.
Additionally, AI models can be trained to predict specific BI-RADS categories directly from images, offering real-time
feedback to radiologists during the assessment. However, integrating AI into BI-RADS evaluation requires careful
calibration to align AI outputs with the standard BI-RADS categories. AI-driven predictions need to be interpretable
and understandable within the context of existing clinical guidelines. Moreover, as BI-RADS relies heavily on visual
assessment, AI’s ability to improve accuracy will be contingent on how well models handle varying image qualities,
patient demographics, and imaging protocols. These factors must be taken into account when developing AI tools for
clinical use to ensure their compatibility with BI-RADS and to enhance clinical decision-making.

C. Challenges of Generalization and the Need for External Valida-tion

A major limitation in deploying AI models in clinical settings is their ability to generalize across different
populations, imaging protocols, and clinical environments. Models trained on one dataset may not perform as
effectively on data from a different hospital or imaging equipment, leading to concerns about their robustness and
reliability. This is particularly critical in breast cancer detection, where variations in imaging devices (e.g.,
mammography vs. digital breast tomosynthesis), scanning techniques, and patient demographics (e.g., age, breast
density, ethnicity) can influence the quality and interpretation of images. Therefore, it is essential that AI models
undergo extensive external validation to ensure their generalizability. This involves testing the models on diverse
 37

datasets that reflect the variety of clinical settings in which the models will be used. External validation is also crucial
for evaluating AI systems’ performance on rare or underrepresented patient subgroups to prevent biases that may
impact diagnostic accuracy. Incorporating multi-center, multi-institutional datasets during the validation phase can
significantly improve the real-world applicability of AI systems, making them more reliable for widespread use.

D. Strategies for Reducing False Positives in AI Breast Cancer Detection

A common challenge in AI-based breast cancer detection is the high incidence of false positives, which can lead to
unnecessary follow-up procedures, patient anxiety, and increased healthcare costs. To reduce false positives, several
strategies have been proposed. First, the use of hybrid models, combining multiple imaging modalities such as
mammography, ultrasound, and MRI, can help increase diagnostic accuracy. Each modality provides complementary
infor-mation, which, when fused, can lead to more robust decision-making. Additionally, integrating clinical data
(e.g., patient history, family history, and BI-RADS assessments) with imaging data can further refine predictions and
reduce the likelihood of false positives. Another promising approach is the application of post-processing techniques,
such as confidence scoring or multi-stage classification models. These methods enable the AI model to assign a
confidence level to each prediction, allowing radiologists to prioritize high-confidence cases and reduce unnecessary
interventions for low-confidence cases. Furthermore, continuous model improvement through retraining on diverse,
high-quality datasets that reflect various patient demograph-ics and imaging conditions is essential for reducing false
positive rates over time.

E. Extremely high accuracy, real-world generalizability, and Dataset Limitations

Multiple studies [76] [77] [71] [72] [73] [74] achieved accuracies exceeding 95-99% on the Wisconsin dataset, yet
none conducted external validation on independent clinical datasets. This pattern of unrealistically high
performance on curated benchmark datasets, without corresponding validation on real-world data, represents a
significant limitation in the current literature. Such findings may mislead clinicians and researchers about the true
clinical readiness of these models and can delay the development of genuinely robust diagnostic systems.
A critical challenge identified in our systematic review is the widespread reliance on small, curated benchmark
datasets that yield inflated performance metrics inconsistent with real-world clinical
38

performance. This ”benchmark dataset problem” manifests in several concerning patterns:

The Wisconsin Dataset Phenomenon: Of the machine learning studies reviewed, several studies used the
Wisconsin Breast Can-cer dataset [76] [77] [78] [71] [72] [73] [74] [19], with reported accuracies ranging from 95%
to 100%. While this dataset enables reproducibility and facilitates algorithm comparison, its use as the primary
validation benchmark is problematic for several reasons. First, the dataset contains only 569 samples, insufficient for
robust evaluation of model generalization, particularly for deep learning ap-proaches. Second, the dataset is artificially
balanced (357 benign, 212 malignant cases), whereas real-world screening populations exhibit cancer prevalence of
only 2-5%, creating substantial class imbalance that models must handle. Third, the dataset consists of pre-extracted
numerical features rather than raw imaging data, bypassing critical challenges of image preprocessing, feature
extraction, and handling variable image quality. Fourth, the dataset explicitly excludes am-biguous or difficult cases,
eliminating precisely the scenarios where AI assistance would be most valuable clinically.
Our analysis reveals that studies using small benchmark datasets consistently report higher accuracies than those
using large, diverse clinical datasets. For instance, studies on the Wisconsin dataset report accuracies of 95-100%
[17][22][23], while the large-scale mammography study by Jungkyu et al. [10], which analyzed nearly 750,000 clinical
data, achieved an AUROC of 0.945, excellent by clinical standards, but notably lower than benchmark dataset results.
This discrepancy suggests that small benchmark datasets fail to capture the complexity, variability, and diagnostic
challenges present in real clinical practice. Factors contributing to this gap include: (1) selection bias in benchmark
datasets toward clear-cut cases; (2) lack of image quality variability seen in routine practice; (3) absence of rare
subtypes and atypical presentations; (4) patient-level overfitting when the same patient contributes multiple samples;
and (5) implicit data leakage through repeated use of the same test sets across multiple studies.
The BreakHis and MIAS Limitations: Similar issues affect histopathological and mammographic datasets.
BreakHis [65] [68] [60], despite containing 7,909 images, includes only 82 patients, creating severe risk of patient-
level overfitting where models may learn patient-specific characteristics rather than generalizable cancer patterns. The
MIAS dataset [75] [49] [62], with only 322 images from 161 patients, suffers from both small sample size and
variable an-

notation quality. Studies reporting >98% accuracy on these datasets

[65] [68] should be interpreted with caution, as these results likely

reflect dataset memorization rather than robust diagnostic capability. Consequences for Clinical Translation: The
reliance on inflated benchmark accuracies has several negative consequences. First, it creates unrealistic expectations
among clinicians and healthcare ad-ministrators about AI system performance, potentially leading to premature
deployment or inappropriate trust in AI recommendations. Second, it diverts research effort toward incremental
improvements on saturated benchmarks rather than addressing real clinical challenges such as handling dense breast
tissue, detecting subtle cancers, or reducing false positives in diverse populations. Third, it makes it difficult to
compare studies meaningfully, as a 99% accuracy on Wisconsin dataset provides little information about clinical
utility compared to a 94% AUROC on a large, diverse screening population. Data imbalance represents one of the
most persistent and signif-icant challenges in developing robust machine learning models for breast cancer
detection. This challenge manifests across multiple dimensions, creating complex obstacles that directly impact
 39

model performance and clinical utility. The fundamental issue stems from the natural prevalence of breast cancer in
screening populations, where malignant cases typically comprise only 2-5% of all mam-mographic examinations. This
severe imbalance between positive (cancerous) and negative (benign or normal) cases creates a learning bias in which
models tend to favor the majority class, leading to high specificity but poor sensitivity. Traditional accuracy metrics
become misleading in such scenarios, as a model achieving 95% accuracy might simply classify all cases as negative
while missing critical cancer cases. Beyond binary classification, the imbalance extends to cancer subtypes such as
invasive ductal carcinoma, lobular carcinoma, and ductal carcinoma in situ, as well as staging classifications. Early-stage
cancers, which are most critical for improving patient outcomes, are often underrepresented in training datasets compared
to more advanced cases. This temporal selection bias arises because advanced cases are more likely to be documented
and available in retrospective datasets, while subtle early-stage presentations remain challenging to collect in sufficient
quantities. Additionally, data imbalance also manifests in demographic representation, with certain populations including
the elderly, minorities, and specific breast density cate-gories being underrepresented in training datasets. Technical
factors compound this issue, as variations in imaging protocols, scanner manufacturers, and institutional practices create
additional sources of data heterogeneity not uniformly distributed across datasets. Var-ious mitigation approaches such
as synthetic minority oversampling techniques (SMOTE) [85], generative adversarial networks (GANs) [86], and focal
loss functions have been explored to address class imbalance. However, each approach introduces limitations; synthetic
data generation may fail to capture the full complexity of real patho-logical presentations, and cost-sensitive learning
approaches require careful calibration to avoid overcorrection that leads to excessive false positives in clinical settings.
The challenge of interpretability presents a critical barrier to clinical adoption, as the ”black box” nature of deep
learning models conflicts with the medical community’s need for transparent and explainable diagnostic reasoning.
Radiologists require understanding of which imaging features and regions influence model predictions to integrate
AI recommendations effectively into their diagnostic workflow. Unlike other AI applications where accuracy
might be the primary concern, medical diagnosis demands transparency in reasoning processes. Clinicians need to
evaluate whether model predictions align with established medical knowledge and whether highlighted features
correspond to known pathological indicators. Furthermore, medical AI systems must satisfy regulatory require-ments
that increasingly emphasize explainability and interpretability. Regulatory frameworks such as the European Union’s
AI Act [87] and FDA guidance stress the importance of algorithmic transparency in high-risk applications like
medical diagnosis. Legal liability concerns arise when diagnostic decisions involve unexplainable AI recom-
mendations, creating additional pressure for interpretable solutions. The acceptance of AI systems by healthcare
professionals directly correlates with their understanding of model behavior. Studies have shown that radiologists are
more likely to trust and effectively utilize AI systems when they can comprehend the basis for predictions.
Conversely, opacity in model reasoning leads either to complete rejection of AI assistance or to dangerous over-
reliance without crit-ical evaluation. Existing explainability methods, including gradient-based techniques such as
GradCAM [88] and LIME [89], attention mechanisms, and feature visualization, provide varying degrees of insight
but remain limited in their clinical utility. These methods often produce explanations that are difficult for clinicians
to interpret or may highlight spurious correlations rather than medically relevant features. The challenge lies in
developing interpretability approaches that align with radiological reasoning patterns and clinical expertise.
Additionally, breast cancer detection requires interpretability at mul-
40

tiple levels: pixel-level feature importance for understanding local tissue characteristics, region-level analysis for
identifying suspicious areas, and global model behavior for understanding overall diagnostic reasoning. Current
approaches struggle to provide coherent explana-tions across these different scales of analysis.
The integration of machine learning and deep learning systems into clinical breast cancer screening workflows
presents multifaceted challenges extending beyond technical performance metrics to prac-tical, operational, and
systemic considerations. Clinical breast cancer screening involves complex workflows with multiple stakeholders,
including technologists, radiologists, medical physicists, and admin-istrative staff [90]. AI systems must seamlessly
integrate into existing Picture Archiving and Communication Systems (PACS), Radiology Information Systems (RIS),
and Electronic Health Records (EHR) without disrupting established procedures. The challenge involves balancing
efficiency gains with the need to maintain quality control and professional oversight. Laboratory performance metrics
often fail to translate directly to clinical environments due to differ-ences in patient populations, imaging
protocols, and operational constraints [91]. Models trained on curated research datasets may exhibit degraded
performance when exposed to the full spectrum of clinical cases, including suboptimal image quality, patient
positioning variations, and equipment differences. Continuous monitoring and validation in clinical settings become
essential but resource-intensive requirements. Furthermore, deep learning models, particularly those processing high-
resolution mammographic images, require significant computational resources that may not be readily available in all
clinical settings [92]. The challenge extends to maintaining model performance while ensuring reasonable processing
times that fit within clinical workflows. Cloud-based solutions introduce additional concerns regarding data privacy,
security, and regulatory compliance, especially given the sensitivity of medical imaging data. Healthcare professionals
require comprehensive training not only on technical system operation but also on understanding AI limitations,
appropri-ate use cases, and integration of AI recommendations with clinical judgment. Resistance to change, varying
levels of technical comfort among staff, and the need for ongoing education present significant implementation
barriers. Developing training programs that address diverse learning needs while maintaining clinical productivity
remains a substantial challenge. Lastly, economic and reimbursement con-siderations add complexity, including initial
implementation costs, ongoing maintenance expenses, potential liability insurance changes, and an uncertain
reimbursement landscape. Healthcare institutions must justify investments in AI technology through demonstrated
improvements in patient outcomes, operational efficiency, or cost reduction, while navigating evolving reimbursement
policies for AI-assisted diagnostic services.

VII. FUTURE DIRECTIONS

The development of explainable artificial intelligence represents a critical frontier in advancing breast cancer
detection systems toward clinically meaningful and trustworthy diagnostic tools [93]. Future research in this domain
must address the fundamental tension between model performance and interpretability while meeting the specific needs
of radiological practice. Clinically aligned explanation methods are essential, focusing on creating explanation
techniques that correspond with radiological reasoning patterns and established medical knowledge. This involves
developing visualization methods that highlight anatomically relevant features such as mass mar-gins, architectural
distortions, and calcification patterns in ways that align with standard radiological assessment criteria like BI-RADS.
Advanced attention mechanisms and concept-based expla-nations that map to medical terminology and diagnostic
 41

criteria will help bridge the gap between computational predictions and clinical understanding. Furthermore, next-
generation explainable AI systems will need to provide coherent explanations across multiple scales of analysis,
ranging from pixel-level feature importance to region-based assessments to global diagnostic reasoning. Hierarchical
explanation frameworks capable of decomposing complex decisions into understandable components while maintaining
consistency across different levels of abstraction will be essential for clinical adoption. Interactive explanation systems
are also a promising direction, en-abling radiologists to query AI models about specific aspects of their reasoning.
These systems could allow exploration of counterfactual scenarios, understanding of uncertainty quantification, and
delivery of explanations tailored to clinical questions. Such interactivity would transform AI from a black-box predictor
into a collaborative diagnos-tic partner. Robust frameworks for evaluating the quality and utility of AI explanations in
clinical contexts are needed, including metrics to assess faithfulness, consistency, and clinical relevance, alongside user
studies involving practicing radiologists to understand how different explanation modalities impact diagnostic
confidence and accuracy. Additionally, future explainable AI systems must be designed to meet regulatory
requirements, incorporating documentation standards, au-dit trails, and explanation stability measures to satisfy
medical device approval processes. Explanation methods should remain consistent across software updates and provide
retrospective explanations for historical decisions.
Federated learning emerges as a transformative approach to ad-dress fundamental challenges of data scarcity,
privacy preservation, and model generalization in breast cancer detection while enabling collaboration across diverse
healthcare institutions [94]–[96]. Fed-erated learning frameworks will enable the development of robust detection
models by leveraging distributed datasets without com-promising patient privacy or violating regulations [97].
Advanced cryptographic techniques such as homomorphic encryption and secure multi-party computation will allow
institutions to collaboratively train models while maintaining complete data sovereignty [98]. These approaches are
especially crucial for rare cancer subtypes and diverse demographic groups that require large-scale collaboration to
achieve sufficient sample sizes. Future federated learning systems will also tackle the challenge of model
generalization across different imaging protocols, equipment manufacturers, and patient populations [99]–[101].
Domain adaptation techniques integrated within federated frameworks will enable models to learn institution-specific
character-istics while extracting generalizable features for cancer detection, re-sulting in more robust models that
perform consistently across diverse clinical environments. Additionally, hierarchical and specialized fed-eration
architectures are expected to develop, where specialized mod-els are built for specific cancer subtypes, breast density
categories, or demographic groups, while maintaining global knowledge sharing [102]. This approach will allow
personalized model adaptation along-side collective learning benefits. Real-time collaborative learning capabilities
will also be implemented, allowing models to adapt and improve continuously as new cases are processed across the
network [103]. Such dynamic learning will facilitate rapid adaptation to emerging imaging technologies, evolving
clinical practices, and novel pathological presentations while preserving privacy and regulatory compliance [104].
Finally, sustainable federated learning networks will require fair incentive mechanisms for participation and robust
quality control measures for distributed data contributions, including reputation systems, contribution assessments,
and automated data quality evaluations to ensure high-quality collaboration [105].
The integration of multiple imaging modalities and clinical data streams represents a promising frontier for
enhancing accuracy, robustness, and clinical utility in breast cancer detection through
42

comprehensive patient assessment [106], [107]. Future multimodal systems will seamlessly integrate mammography,
digital breast to-mosynthesis (DBT) [108], breast MRI [109], and breast ultrasound data to provide comprehensive
diagnostic assessments. Deep learning architectures tailored for multimodal fusion will learn complementary features
across imaging modalities: mammography offers population screening capabilities [110], DBT improves visualization
of overlap-ping tissues [111], MRI delivers enhanced soft tissue contrast [112], [113], and ultrasound provides real-
time assessment. These integrated approaches are expected to significantly improve sensitivity for detecting cancers
that may be subtle or invisible in single modalities. Beyond imaging, sophisticated fusion frameworks will incorporate
diverse clinical data including family history, genetic markers such as BRCA1/BRCA2 status, previous imaging
results, demographic factors, and patient-reported symptoms. Advanced attention mech-anisms [114]–[119] and graph
neural networks [120]–[125] will model complex relationships between clinical variables and imaging features,
enabling personalized risk assessment and tailored screening recommendations. Longitudinal data integration will also
play a key role, with temporal fusion models analyzing changes across multiple imaging sessions to detect subtle
progression patterns indicative of early malignancy. This temporal data fusion could identify growth patterns,
morphological changes, and emerging features that are only apparent through time-series analysis, potentially enabling
earlier detection of developing cancers [126]. Cross-modal consistency and validation mechanisms will improve
prediction confidence by leveraging agreement across modalities, while discrepancies can be flagged for further review
or additional imaging. Adaptive fusion strategies [127]–[129] will weight data sources dynamically based on patient
characteristics, image quality, and clinical context, optimizing diagnostic accuracy and potentially reducing
unnecessary imaging [130]. Future research will also address practical implementation challenges including
standardization of multimodal data formats, development of efficient computational architectures for real-time
processing [131], and creation of user interfaces that effectively present integrated multimodal results to clinicians in
actionable formats. Visualization tools that assist radiologists in understanding and interpreting complex multimodal
predictions while maintaining clinical workflow efficiency will be essential for successful real-world deployment [132].

REFERENCES

[1] N. Manjunathan, N. Gomathi, and S. Muthulingam, “Early detection of breast cancer using machine
learning,” in 2023 International Con-ference on Sustainable Computing and Smart Systems (ICSCSS),
2023,
pp. 165–169. doi: 10.1109/ICSCSS57650.2023.10169777 .
[2] S. Pawar, P. Bagal, P. Shukla, and A. Dawkhar, “Detection of breast cancer using machine learning
classifier,” in 2021 Asian Confer-ence on Innovation in Technology (ASIANCON), 2021, pp. 1–5. doi:
10.1109/ASIANCON51346.2021.9544767 .
[3] N. Khuriwal and N. Mishra, “Breast cancer diagnosis using adap-tive voting ensemble machine learning
algorithm,” in 2018 IEEMA Engineer Infinite Conference (eTechNxT), 2018, pp. 1–5. doi:
10.1109/ETECHNXT.2018.8385355 .
[4] C. Bista, A. M, S. Slimanzay, M. S. Sheikh, and P. Srinivasa Rao, “Breast cancer prediction system
 43

utilizing machine learning algo-rithms,” in 2024 IEEE AITU: Digital Generation, 2024, pp. 80–84. doi:
10.1109/IEEECONF61558.2024.10585589 .
[5] V. Chauhan, R. Shukla, S. Ansari, A. Bindiya, and M. K. Gupta, “A systematic survey on breast
cancer detection using machine learning,” in 2023 6th International Conference on Contemporary
Computing and Informatics (IC3I), vol. 6, 2023, pp. 30–35. doi: 10.1109/IC3I59117.2023.10397976 .

[6] A. Tale, A. Barhate, P. Verma, and P. Gourshettiwar, “A review of machine learning algorithms for
detection of breast cancer,” in 2024 4th International Conference on Ubiquitous Computing and Intelligent
Information Systems (ICUIS), 2024, pp. 1148–1153. doi: 10.1109/ICUIS64676.2024.10866542 .
[7] J. Mu, A. Nazar, M. A. Ali, and A. Hussain, “Integrating machine learning with omics data for early
detection in breast cancer,” Gene Reports, vol. 41, p. 102325, Dec. 2025. doi: 10.1016/[Link].2025.102325 .
[Online]. Available: [Link] 10.1016/[Link].2025.102325
[8] H. L. P, S. Ravi, and S. A, “Breast cancer detection using machine learning in medical imaging – a survey,”
Procedia Computer Science, vol. 239, p. 2235–2242, 2024. doi: 10.1016/[Link].2024.06.414 .
[Online]. Available: [Link]
[9] A. S. Boddu and A. Jan, “A systematic review of machine learning algorithms for breast cancer detection,”
Tissue and Cell, vol. 95,
p. 102929, Aug. 2025. doi: 10.1016/[Link].2025.102929 . [Online].
Available: [Link]
[10] J. Park, J. Witowski, Y. Xu, H. Trivedi, J. Gichoya, B. Brown-Mulry,
M. Westerhoff, L. Moy, L. Heacock, A. Lewin, and K. J. Geras, “A multi-modal ai system for screening
mammography: Integrating 2d and 3d imaging to improve breast cancer detection in a prospective clinical
study,” 2025. [Online]. Available: [Link]
[11] M. M. J. K. C. A. A. N. . M. A. S. Proshenjit Sarker, Amel Ksibi, “Breast cancer prediction with feature-
selected xgb classifier, optimized by metaheuristic algorithms,” Journal of Big Data, vol. 12, p. 78, 2025. doi:
10.1186/s40537-025-01132-7 .
[12] T. M. Chowdhury and A. R. M. Kamal, “An efficient and inter-pretable machine learning model for
classifying breast cancer subtypes using gene expression profiles,” Engineering, Technology & Applied Science
Research, vol. 15, no. 4, pp. 24 196–24 203, 2025. doi: 10.48084/etasr.11179 .
[13] H. Li, C. T. Zhang, H. G. Shao et al., “Prediction models of breast cancer molecular subtypes based on
multimodal ultrasound and clinical features,” BMC Cancer, vol. 25, p. 886, 2025. doi: 10.1186/s12885-025-
14233-6 .
[14] A. M. Hassan, S. M. Naeem, M. A. A. Eldosoky et al., “Multi-omics-based machine learning for the
subtype classification of breast cancer,” Arab Journal of Science and Engineering, vol. 50, pp. 1339–1352,
2025. doi: 10.1007/s13369-024-09341-7 . [Online]. Available: [Link]
[15] D. Zuo, L. Yang, Y. Jin et al., “Machine learning-based models for the prediction of breast cancer
recurrence risk,” BMC Medical Informatics and Decision Making, vol. 23, p. 276, 2023. doi: 10.1186/s12911-
023-02377-z . [Online]. Available: [Link] 1186/s12911-023-02377-z
 44

[16] A. A. Nahid, M. J. Raihan, and A. A. M. Bulbul, “Breast cancer classification along with feature
prioritization using machine learning algorithms,” Health Technology, vol. 12, pp. 1061–1069, 2022. doi:
10.1007/s12553-022-00710-6 . [Online]. Available: [Link]
[17] S. S. Kumar, V. Kathiresan, R. Gomathi, R. Karthik, S. S. Priya, and C. N. Jasmine, “Advancements in
breast cancer detection using machine learning techniques for early and accurate prediction,” in 2024 Second
International Conference on Intelligent Cyber Physical Systems and Internet of Things (ICoICI), 2024, pp.
730–737. doi: 10.1109/ICoICI62503.2024.10695994 .
[18] H. Al-Asria, H. Mousannif, H. Al Moatassime, and T. Noel, “Breast cancer detection using machine learning
approaches: A comparative study,” International Journal of Electrical and Computer Engineering, vol. 13, no.
1, pp. 736–745, 2023.
[19] P. Karatza, K. V. Dalakleidi, M. Athanasiou, and K. S. Nikita, “Inter-pretability methods of machine learning
algorithms with applications in breast cancer diagnosis,” arXiv preprint, 2022, arXiv:2202.02131.
[20] T. Islam, M. A. Sheakh, M. S. Tahosin, M. H. Hena, S. Akash, Y. A. Bin Jardan, G.
FentahunWondmie, H.-A. Nafidi, and
M. Bourhia, “Predictive modeling for breast cancer classification in the context of bangladeshi patients by
use of machine learning approach with explainable ai,” Scientific Reports, vol. 14, no. 1, Apr. 2024. doi:
10.1038/s41598-024-57740-5 . [Online]. Available: [Link]
[21] T. Arravalli, K. Chadaga, H. Muralikrishna, N. Sampathila, D. Cenitta,
R. Chadaga, and K. S. Swathi, “Detection of breast cancer using machine learning and explainable artificial
intelligence,” Scientific Reports, vol. 15, no. 1, Jul. 2025. doi: 10.1038/s41598-025-12644-w . [Online].
Available: [Link]
[22] K. Puttegowda, V. V, H. Ranjan Kumar, K. Sudheesh, K. Prabhavathi,
R. Vinayakumar, and K. Tabianan, “Enhanced machine learning models for accurate breast cancer
mammogram classification,” Global Transitions, vol. 7, p. 276–295, 2025. doi: 10.1016/[Link].2025.04.007 .
[Online]. Available: [Link]
[23] A. N. Nuzla, A. K. M. Nabeel, W. A. S. Nirmal, P. B. Hewavithana, and M. L. Jayatilake, “Machine
learning-based classification of histological subtypes of invasive breast cancer using mri contralateral
breast texture features,” Scientific Reports, vol. 15, no. 1, Nov. 2025. doi: 10.1038/s41598-025-23498-
7 . [Online]. Available: http:
//[Link]/10.1038/s41598-025-23498-7
[24] E. M. Tegaw and B. B. Asfaw, “Machine learning analysis of survival outcomes in breast cancer patients
treated with chemotherapy, hormone therapy, surgery, and radiotherapy,” Scientific Reports, vol. 15, no. 1,
Jul. 2025. doi: 10.1038/s41598-025-97763-0 . [Online]. Available: [Link]
97763-0
[25] S. Kaindal and B. Venkataramana, “A comparative analysis of parametric survival models and machine
learning methods in breast cancer prognosis,” Scientific Reports, vol. 15, no. 1, Aug. 2025. doi:
10.1038/s41598-025-15696-0 . [Online]. Available: [Link]
[26] Y. Zhang, W. An, C. Wang, X. Liu, Q. Zhang, Y. Zhang, and
 45

S. Cheng, “Novel models based on machine learning to predict the prognosis of metaplastic breast
cancer,” The Breast, vol. 79, p. 103858, Feb. 2025. doi: 10.1016/[Link].2024.103858 . [Online]. Available:
[Link]
F. F. Miglietta, A. Collesei, C. Vernieri, T. Giarratano, C. Giorgi, Girardi, G. Griguolo, M. Cacciatore, A.
Botticelli, A. Vingiani, Fotia, F. Piacentini, D. Massa, F. Zangh`ı, M. Marino, G. Pruneri,
M. Fassan, A. Dei Tos, M. Dieci, and V. Guarneri, “Development of two machine learning models to
predict conversion from primary her2-0 breast cancer to her2-low metastases: a proof-of-concept study,”
ESMO Open, vol. 10, no. 1, p. 104087, Jan. 2025. doi: 10.1016/[Link].2024.104087 . [Online].
Available: [Link]
[27] M. Koume, L. Seguin, J. Mancini, M.-K. Bendiane, A.-D. Bouhnik, and R. Urena, “Predicting fear of
breast cancer recurrence in women five years after diagnosis using machine learning and healthcare
reimbursement data from the french nationwide vican survey,” International Journal of Medical
Informatics, vol. 193, p. 105705, Jan. 2025. doi: 10.1016/[Link].2024.105705 . [Online]. Available:
[Link]
[28] C. Aroef, Y. Rivan, and Z. Rustam, “Comparing random forest and support vector machines for breast
cancer classification,” TELKOM-NIKA, vol. 18, no. 2, pp. xxx–xxx, 2024. doi: 10.12928/telkom-
nika.v18i2.14785 .
[29] A. La Moglia and K. Mohamad Almustafa, “Breast cancer prediction using machine learning classification
algorithms,” Intelligence-Based Medicine, vol. 11, p. 100193, 2025. doi: 10.1016/[Link].2024.100193 .
[Online]. Available: [Link]
[30] V. Priyadarshni, S. K. Sharma, M. K. I. Rahmani, B. Kaushik, and R. Almajalid, “Machine learning
techniques using deep instinctive encoder-based feature extraction for optimized breast cancer
detection,” Computers, Materials and Continua, vol. 78, no. 2, p. 2441–2468, 2024. doi:
10.32604/cmc.2024.044963 . [Online].
Available: [Link]
[31] J. Iyadurai, M. Chandrasekharan, S. Muthusamy, and H. Panchal, “An extensive review on emerging
advancements in thermography and convolutional neural networks for breast cancer detection,” Wireless
Personal Communications, vol. 137, no. 3, p. 1797–1821, Jul. 2024. doi: 10.1007/s11277-024-11466-9 .
[Online]. Available: [Link]
[32] U. R. Acharya, E. Y. K. Ng, J.-H. Tan, and S. V. Sree, “Thermography based breast cancer detection
using texture features and support vector machine,” Journal of Medical Systems, vol. 36, no. 3, p.
1503–1510, Oct. 2010. doi: 10.1007/s10916-010-9611-z . [Online].
Available: [Link]
[33] N. P. Dharani, I. G. Immadi, and M. V. Narayana, “Enhanced deep learning model for diagnosing
breast cancer using thermal images,” Soft Computing, vol. 28, pp. 8423–8434, 2024. doi: 10.1007/s00500-
024-09742-8 . [Online]. Available: [Link] 1007/s00500-024-09742-8
[34] A. Y. A. B. Ahmad, J. A. Alzubi, M. Vasanthan et al., “Efficient hybrid heuristic adopted deep learning
framework for diagnosing breast cancer using thermography images,” Scientific Reports, vol. 15, p. 13605,
46

2025. doi: 10.1038/s41598-025-96827-5 . [Online]. Available:

[Link]

[35] N. Aidossov, V. Zarikas, A. Mashekova, Y. Zhao, E. Y. K. Ng, A. Midlenko, and O. Mukhmetov,

“Evaluation of integrated cnn, transfer learning, and bn with thermography for breast cancer detection,”
Applied Sciences, vol. 13, no. 1, p. 600, Jan. 2023. doi: 10.3390/app13010600 . [Online]. Available:
http:
//[Link]/10.3390/app13010600
[36] A. Alshehri and D. AlSaeed, “Breast cancer detection in thermography using convolutional neural networks
(cnns) with deep attention mechanisms,” Applied Sciences, vol. 12, no. 24, p. 12922, Dec. 2022. doi:
10.3390/app122412922 . [Online]. Available: [Link] 3390/app122412922
[37] M. A. S. Al Husaini, M. H. Habaebi, and M. R. Islam, “Real-time thermography for breast cancer
detection with deep learning,” Discover Artificial Intelligence, vol. 4, no. 1, Aug. 2024. doi:
10.1007/s44163-024-00157-w . [Online]. Available: http:
//[Link]/10.1007/s44163-024-00157-w
[38] R. Jalloul, C. H. Krishnappa, V. I. Agughasi, and R. Alkhatib, “Enhancing early breast cancer detection with
infrared thermography: A comparative evaluation of deep learning and machine learning models,”
Technologies, vol. 13, no. 1, p. 7, Dec. 2024. doi: 10.3390/technologies13010007 . [Online]. Available:
[Link] 10.3390/technologies13010007
[39] O. Attallah, “Harnessing infrared thermography and multi-convolutional neural networks for early breast cancer
detection,” Scien-tific Reports, vol. 15, no. 1, Jul. 2025. doi: 10.1038/s41598-025-09330-
2 . [Online]. Available: [Link]
[40] A. A. Khan and A. S. Arora, “Thermography as an economical alternative modality to mammography for
early detection of breast cancer,” Journal of Healthcare Engineering, vol. 2021, p. 1–8, Jul. 2021. doi:
10.1155/2021/5543101 . [Online]. Available: [Link]
[41] A. Mehmood, Y. Hu, and S. H. Khan, “A novel channel boosted residual cnn-transformer with regional-
boundary learning for breast cancer detection,” 2025. [Online]. Available: [Link]
[42] M. Fatima, M. A. Khan, A. M. Mirza et al., “An adaptive deep learning approach based on inbnfus and
cnnden-gru networks for breast cancer and maternal fetal classification using ultrasound images,” Scientific
Reports, vol. 15, no. 1, p. 22027, 2025. doi: 10.1038/s41598-025-03402-z . [Online]. Available:
[Link] 1038/s41598-025-03402-z
[43] J. Chen, T. Pan, Z. Zhu et al., “A deep learning-based multimodal medical imaging model for breast cancer
screening,” Scientific Reports, vol. 15, no. 1, p. 14696, 2025. doi: 10.1038/s41598-025-99535-2 .
[Online]. Available: [Link]
[44] M. M. Labonno, D. M. Asadujjaman, M. M. Rahman, A. Tamim,
M. J. Ferdous, and R. M. Mahi, “Early detection and classification of breast cancer using deep learning
techniques,” 2025. [Online].
Available: [Link]
[45] A. Bunnell, D. Valdez, T. K. Wolfgruber, B. Quon, K. Hung, B. Y. Hernandez, T. B. Seto, J. Killeen, M.
 47

Miyoshi, P. Sadowski, and

J. A. Shepherd, “Deep learning predicts mammographic breast density in clinical breast ultrasound images,”
2024. [Online]. Available: [Link]
[46] I. Chamveha, S. Chaiyungyuen, S. Worakriangkrai, N. Prasawang, W. Chaisangmongkon, P. Korpraphong,
V. Suvannarerg, S. Thiravit, C. Kannawat, K. Rungsinaporn, S. Issaragrisil, P. Chadbunchachai, P.
Gatechumpol, C. Muktabhant, and P. Sereerat, “Deep learning-based breast cancer detection in
mammography: A multi-center validation study in thai population,” 2025. [Online]. Available:
[Link]
[47] O. Camurdan, T. Tanyel, E. A. Cerekci et al., “Annotation-efficient, patch-based, explainable deep learning
using curriculum method for breast cancer detection in screening mammography,” Insights into Imaging, vol.
16, p. 60, 2025. doi: 10.1186/s13244-025-01922-w .
[Online]. Available: [Link]
[48] A. Abdel-Wahab, V. Pulipati, S. Joshi et al., “Deep learning framework for breast cancer detection and
segmentation using efficientnet and u-net with hyperparameter optimization,” SN Computer Science, vol. 6,
p. 528, 2025. doi: 10.1007/s42979-025-03900-x . [Online]. Available: [Link]
03900-x
[49] J. W. Seo, Y. J. Kim, and K. G. Kim, “Leveraging paired mammogram views with deep learning for
comprehensive breast cancer detection,” Scientific Reports, vol. 15, no. 1, p. 4406, 2025. doi:
10.1038/s41598-025-88907-3 . [Online]. Available: https:
//[Link]/10.1038/s41598-025-88907-3
[50] S. Zeng, H. Chen, R. Jing et al., “An assessment of breast cancer her2, er, and pr expressions based on
mammography using deep learning with convolutional neural networks,” Scientific Reports, vol. 15, no. 1,
p. 4826, 2025. doi: 10.1038/s41598-024-83597-9 . [Online]. Available: [Link]
83597-9

[51] P. S. R. C. Murty, C. Anuradha, P. A. Naidu et al., “Integrative hybrid deep learning for
enhanced breast cancer diagnosis: leveraging the wisconsin breast cancer database and the cbis-
ddsm dataset,” Scientific Reports, vol. 14, no. 1, p. 26287, 2024. doi: 10.1038/s41598-024-74305-8 .
[Online]. Available:
[Link]
[52] N. Chaudhary and Z. Dhunny, “An artificial intelligence model for early stage breast cancer detection
from biopsy images,” 2025. [Online]. Available: [Link]
[53] M. Tafavvoghi, L. A. Bongo, A. B. Delgado, N. Shvetsov, A. Sildnes,
L. Moi, L.-T. R. Busund, and K. Møllersen, “Automating tumor-infiltrating lymphocyte assessment in breast
cancer histopathology images using qupath: a transparent and accessible machine learning pipeline,” 2025.
[Online]. Available: [Link]
[54] R. O. Ogundokun, P. A. Owolawi, and C. Tu, “Optimized deep feature learning with hybrid
ensemble soft voting for early breast cancer histopathological image classification,” Computers,
Materials and Continua, vol. 84, no. 3, p. 4869–4885, 2025. doi: 10.32604/cmc.2025.064944 . [Online].
48

Available: [Link] 32604/cmc.2025.064944

[55] F. Li, J. Ma, T. Wen, Z. Tian, and H.-N. Liang, “Hi-net: A novel histopathologic image segmentation
model for metastatic breast cancer via lightweight dataset construction,” Heliyon, vol. 10, no. 19,
p. e38410, Oct. 2024. doi: 10.1016/[Link].2024.e38410 . [Online].
Available: [Link]
[56] J. Yang, J. Ju, L. Guo, B. Ji, S. Shi, Z. Yang, S. Gao, X. Yuan,
G. Tian, Y. Liang, and P. Yuan, “Prediction of her2-positive breast cancer recurrence and metastasis risk
from histopathological images and clinical information via multimodal deep learning,” Computational and
Structural Biotechnology Journal, vol. 20, p. 333–342, 2022. doi: 10.1016/[Link].2021.12.028 . [Online].
Available: [Link]
[57] A. Rehman, M. Mujahid, R. Damasevicius, F. S. Alamri, and
T. Saba, “Densely convolutional bu-net framework for breast multi-organ cancer nuclei segmentation through
histopathological slides and classification using optimized features,” Computer Modeling in Engineering and
Sciences, vol. 141, no. 3, p. 2375–2397, 2024. doi: 10.32604/cmes.2024.056937 . [Online]. Available: http:
//[Link]/10.32604/cmes.2024.056937
[58] R. Ochoa-Ornelas, A. Gudin˜o-Ochoa, J. A. Garc´ıa-Rodr´ıguez, and S. Uribe-Toscano, “A robust
transfer learning approach with histopathological images for lung and colon cancer detection using
efficientnetb3,” Healthcare Analytics, vol. 7, p. 100391, Jun. 2025. doi: 10.1016/[Link].2025.100391 .
[Online]. Available: [Link]
[59] M. Pradeepa, B. Sharmila, and M. Nirmala, “A hybrid deep learning model efficientnet with gru for
breast cancer detection from histopathology images,” Scientific Reports, vol. 15, p. 24633, 2025. doi:
10.1038/s41598-025-00930-6 . [Online]. Available: https:
//[Link]/10.1038/s41598-025-00930-6
[60] T. C. Phan and H. N. Huu, “Enhancing breast cancer detection with advanced deep learning techniques
for ki-67 nuclear protein analysis,” SN Computer Science, vol. 5, p. 663, 2024. doi: 10.1007/s42979-024-
03004-y . [Online]. Available: [Link] 1007/s42979-024-03004-y
[61] S. Vijayalakshmi, B. K. Pandey, D. Pandey et al., “Innovative deep learning classifiers for breast
cancer detection through hybrid feature extraction techniques,” Scientific Reports, vol. 15, p. 22212,
2025. doi: 10.1038/s41598-025-06669-4 . [Online]. Available:
[Link]
[62] S. R. S. Chakravarthy, N. Bharanidharan, V. V. Kumar et al., “Deep transfer learning with fuzzy
ensemble approach for the early detection of breast cancer,” BMC Medical Imaging, vol. 24,
p. 82, 2024. doi: 10.1186/s12880-024-01267-8 . [Online]. Available: [Link]
[63] S. A. Alzakari, S. Hassairi, A. A. A. Hussan et al., “Enhancing breast cancer classification using a
deep sparse wavelet autoencoder approach,” Scientific Reports, vol. 15, no. 1, p. 26194, 2025. doi:
10.1038/s41598-025-11816-y . [Online]. Available: [Link] 1038/s41598-025-11816-y
[64] N. Sengodan, “Breast cancer histopathology classification using cbam-efficientnetv2 with transfer
learning,” 2025. [Online]. Available: [Link]
 49

[65] S. K. Singh and K. S. Patnaik, “Patho-ai: A perceptive breast cancer identification and classification using
deep learning methods integrated with explainable ai,” SN Computer Science, vol. 6, p. 619, 2025. doi:
10.1007/s42979-025-04170-3 . [Online]. Available: [Link]
[66] L. S. Potti and S. Maruthuperumal, “Breast cancer cell detec-tion using fcm and prediction using unet
based deep convo-lutional neural network,” in 2024 5th IEEE Global Conference for Advancement in
Technology (GCAT), 2024, pp. 1–6. doi: 10.1109/GCAT62922.2024.10924106 .
[67] A. K. Singh, B. D. Mazumdar, R. Raja et al., “Transforming early breast cancer detection: A deep learning
approach using convolutional neural networks and advanced classification techniques,” International Journal
of Computational Intelligence Systems, vol. 18,
p. 134, 2025. doi: 10.1007/s44196-025-00876-7 . [Online]. Available: [Link]
00876-7
[68] M. Khan, M. M. Su’ud, M. M. Alam et al., “Enhancing breast cancer detection through optimized thermal
image analysis using prms-net deep learning approach,” Journal of Digital Imaging: Informatics in Medicine,
2025. doi: 10.1007/s10278-025-01465-y . [Online]. Available: [Link]
[69] A. Sharma and P. K. Mishra, “Inception unet architecture for breast tumor segmentation and detection
using hybrid deep learning approach,” Multimedia Tools and Applications, vol. 84, pp. 28 225–28 263, 2025.
doi: 10.1007/s11042-024-20234-7 . [Online]. Available:
[Link]
[70] I. Koc, W. Tashan, I. Shayea, and A. Zhetpisbayeva, “Breast cancer detection based on machine learning,” in
2024 IEEE 13th International Conference on Communication Systems and Network Technologies (CSNT),
2024, pp. 1–6. doi: 10.1109/CSNT60213.2024.10545785 .
[71] H. Sharma, P. Singh, and A. Bhardwaj, “Breast cancer detection: Com-parative analysis of machine learning
classification techniques,” in 2022 International Conference on Emerging Smart Computing and Informat-ics
(ESCI), 2022, pp. 1–6. doi: 10.1109/ESCI53509.2022.9758188 .
[72] S. Chen, “Diagnosis of breast cancer using random forests,” Procedia Computer Science, vol. 218, pp. 429–
437, 2023. doi: 10.1016/[Link].2023.01.025 .
[73] ——, “Breast cancer prediction based on rf-svm,” in Applied and Com-putational Engineering, vol. 8, 2023, pp.
657–666. doi: 10.54254/2755-2721/8/20230293 .
[74] D. Ghadge, S. Hon, T. Saraf, T. Wagh, A. Tambe, and Y. Desh-mukh, “Analysis on machine learning-
based early breast cancer de-tection,” in 2024 4th International Conference on Innovative Prac-tices in
Technology and Management (ICIPTM), 2024, pp. 1–5. doi: 10.1109/ICIPTM59628.2024.10563587 .
[75] A. Rovshenov and S. Peker, “Performance comparison of different machine learning techniques for early
prediction of breast cancer using wisconsin breast cancer dataset,” in 2022 3rd International Informatics and
Software Engineering Conference (IISEC), 2022, pp. 1–6. doi: 10.1109/IISEC56263.2022.9998248 .
[76] M. Gupta and B. Gupta, “A comparative study of breast cancer diagnosis using supervised machine learning
techniques,” in 2018 Second International Conference on Computing Methodologies and Communication
(ICCMC), 2018, pp. 997–1002. doi: 10.1109/IC-CMC.2018.8487537 .
[77] M. A. Haque, M. Majumder, and S. M. S. Islam, “Early screen-ing of breast cancer using machine learning
50

algorithms: A com-parative study,” in 2024 IEEE International Conference on Future Machine Learning and
Data Science (FMLDS), 2024, pp. 51–56. doi: 10.1109/FMLDS63805.2024.00019 .
[78] S. N. Duggan, M. Azharuddin, R. Herna´ndez, C. Robertson,
D. Cooper, E. McCall, P. Manson, G. Virgili, M. Clarke, S. Treweek, and M. Brazzelli, “Supplemental
imaging modalities for breast cancer screening in women with dense breasts: A systematic review with
economic considerations,” The Breast, vol. 85, p. 104668, Feb. 2026. doi: 10.1016/[Link].2025.104668 .
[Online]. Available: [Link]
[79] A. Mathur, A. M. Dudhiyawala, S. R. Dey, and S. Saha, “Toward accurate breast cancer classification: A
review of multi-modal machine learning approaches,” Methods, vol. 246, p. 48–61, Feb. 2026. doi:
10.1016/[Link].2025.10.011 . [Online]. Available: [Link]
[80] H. Figueiras, J. Domingues, N. Matela, and N. Garcia, “Self-supervised learning for breast cancer detection:
A review,” Computers in Biology and Medicine, vol. 198, p. 111245, Nov. 2025. doi:
10.1016/[Link].2025.111245 . [Online]. Available: http:
//[Link]/10.1016/[Link].2025.111245

[81] X. J. Tan, W. L. Cheor, E. M. Cheng, C. C. Lim, and K. S. A. Rahman, “Generative artificial intelligence (gai)
in breast cancer diagnosis and treatment: A systematic review,” Computers, Materials & Continua, vol. 84,
no. 2, p. 2015–2060, 2025. doi: 10.32604/cmc.2025.063407 .
[Online]. Available: [Link]
[82] N. Shifa, M. Saleh, Y. Akbari, and S. Al Maadeed, “A review of explainable ai techniques and their
evaluation in mammography for breast cancer screening,” Clinical Imaging, vol. 123, p. 110492, Jul.
2025. doi: 10.1016/[Link].2025.110492 . [Online]. Available:
[Link]
[83] I. Haq, Z. Gong, H. Liang, W. Zhang, R. Khan, L. Gu, R. Eils, Y. Kang, and B. Huang, “A review of breast
cancer histopathology image analysis with deep learning: Challenges, innovations, and clinical integration,”
Image and Vision Computing, vol. 162, p. 105708, Oct. 2025. doi: 10.1016/[Link].2025.105708 . [Online].
Available: [Link]
[84] I. Chhillar and A. Singh, “An improved soft voting-based machine learning technique to detect breast
cancer utilizing effective feature selection and smote-enn class balancing,” Discover Artificial
Intelligence, vol. 5, no. 1, Jan. 2025. doi: 10.1007/s44163-025-00224-w . [Online]. Available:
[Link]
[85] F. Gurcan and A. Soylu, “Synthetic boosted resampling using deep generative adversarial networks: A
novel approach to improve cancer prediction from imbalanced datasets,” Cancers, vol. 16, no. 23, p. 4046,
Dec. 2024. doi: 10.3390/cancers16234046 . [Online]. Available: [Link]
[86] R. Upreti, P. G. Lind, A. Elmokashfi, and A. Yazidi, “Trustworthy machine learning in the context of
security and privacy,” International Journal of Information Security, vol. 23, no. 3, p. 2287–2314,
Apr. 2024. doi: 10.1007/s10207-024-00813-3 . [Online]. Available: [Link]
00813-3
[87] M. Dutta, K. M. Mehedi Hasan, A. Akter, M. H. Rahman, and M. Assaduzzaman, “An interpretable
 51

machine learning-based breast cancer classification using xgboost, shap, and lime,” Bulletin of Electrical
Engineering and Informatics, vol. 13, no. 6, Dec. 2024. doi: 10.11591/eei.v13i6.7866 . [Online]. Available:
[Link] 10.11591/eei.v13i6.7866
[88] B. U. Maheswari, A. A, A. Avvaru, A. Tandon, and R. P. de Prado, “Interpretable machine learning model
for breast cancer prediction using lime and shap,” in 2024 IEEE 9th International Confer-ence for
Convergence in Technology (I2CT), 2024, pp. 1–6. doi: 10.1109/I2CT61223.2024.10543965 .
[89] U. A. Usmani, A. Y. Usmani, and M. U. Usmani, “Ensuring trustworthy machine learning: Ethical
foundations, robust algorithms, and respon-sible applications,” in 2023 International Conference on
Computing, Communication, and Intelligent Systems (ICCCIS). IEEE, 2023, pp. 576–583.
[90] F. Cheng, D. Liu, F. Du, Y. Lin, A. Zytek, H. Li, H. Qu, and
K. Veeramachaneni, “Vbridge: Connecting the dots between features, explanations, and data for healthcare
models,” CoRR, 2021.
[91] U. Bhatt, “Trustworthy machine learning: From algorithmic trans-parency to decision support,” Ph.D.
dissertation, 2024.
[92] R. Gonzales Martinez and D.-M. van Dongen, “Deep learning algorithms for the early detection of breast
cancer: A comparative study with traditional machine learning,” Informatics in Medicine Unlocked, vol.
41, p. 101317, 2023. doi: 10.1016/[Link].2023.101317 . [Online]. Available:
[Link]
[93] S. Kumbhare, A. [Link], and S. Shinde, “Federated learning aided breast cancer detection with intelligent
heuristic-based deep learning framework,” Biomedical Signal Processing and Control, vol. 86, p. 105080,
Sep. 2023. doi: 10.1016/[Link].2023.105080 . [Online].
Available: [Link]
[94] B. Karthiga, K. Praneeth, V. Saravanan, and T. Rama Krishna Rao, “Enhancing cancer detection in medical
imaging through federated learning and explainable artificial intelligence: A hybrid approach for optimized
diagnostics,” Egyptian Informatics Journal, vol. 31, p. 100751, Sep. 2025. doi: 10.1016/[Link].2025.100751 .
[Online]. Available: [Link]
[95] J. B. Awotunde, C. O. Abikoye, B. Brahma, E. T. Oladipupo, and A. Bandyopadhyay, “Federated
learning augmented with convolutional neural networks for brain cancer classification,” Procedia
Computer Science, vol. 258, p. 2617–2626, 2025. doi: 10.1016/[Link].2025.04.523 . [Online].
Available: [Link]
[96] S. O. Folorunso, J. B. Awotunde, A. A. Kawu, and O. Banjo, Performance evaluation of federated learning
algorithms using breast cancer dataset. Elsevier, 2024, p. 95–114. doi: 10.1016/b978-0-443-13897-
3.00015-1 . [Online]. Available: [Link] B978-0-443-13897-3.00015-1
[97] A. Rehman, H. Xing, L. Feng, M. Hussain, N. Gulzar, M. Adnan Khan, A. Hussain, and D. Saeed, “Fedcscd-
gan: A secure and collaborative framework for clinical cancer diagnosis via optimized federated learning and
gan,” Biomedical Signal Processing and Control, vol. 89, p. 105893, Mar. 2024. doi:
10.1016/[Link].2023.105893 . [Online]. Available: [Link]
[98] R. Joshi, S. Kumari, K. Pandey, and R. Badola, A collaborated federated learning for healthcare informatics:
52

solution and challenges. Elsevier, 2026, p. 171–188. doi: 10.1016/b978-0-443-33789-5.00024-8 . [Online].

Available: [Link]
[99] H.-R. Zhang, K.-X. Wang, X.-Y. Liang, and Y.-F. Yu, “Dups: Data poisoning attacks with uncertain
sample selection for federated learning,” Computer Networks, vol. 256, p. 110909, Jan. 2025. doi:
10.1016/[Link].2024.110909 . [Online]. Available: [Link] 10.1016/[Link].2024.110909
[100] I. Z. Yao, M. Dong, and W. Y. Hwang, “Deep learning applications in clinical cancer detection: A review of
implementation challenges and solutions,” Mayo Clinic Proceedings: Digital Health, vol. 3, no. 3, p.
100253, Sep. 2025. doi: 10.1016/[Link].2025.100253 . [Online]. Available:
[Link]
[101] S. Panda, R. Dubey, B. Jena, V. Pareek, L.-W. Tsai, and S. Saxena, “Federated learning frameworks in smart e-
healthcare: A systematic literature review with bias evaluation,” Applied Soft Computing, vol. 171, p. 112747,
Mar. 2025. doi: 10.1016/[Link].2025.112747 . [Online]. Available: [Link]
[102] D. Upreti, E. Yang, H. Kim, and C. Seo, “A comprehensive survey on federated learning in the healthcare area:
Concept and applications,” Computer Modeling in Engineering and Sciences, vol. 140, no. 3, p. 2239–
2274, 2024. doi: 10.32604/cmes.2024.048932 . [Online]. Available:
[Link]
[103] H. Li, C. Li, J. Wang, A. Yang, Z. Ma, Z. Zhang, and D. Hua, “Review on security of federated learning
and its application in healthcare,” Future Generation Computer Systems, vol. 144, p. 271–290, Jul.
2023. doi: 10.1016/[Link].2023.02.021 . [Online]. Available: [Link]
[104] M. Narula, J. Meena, and D. K. Vishwakarma, “A comprehensive review on federated learning for data-
sensitive application: Open issues and challenges,” Engineering Applications of Artificial Intelligence, vol.
133, p. 108128, Jul. 2024. doi: 10.1016/[Link].2024.108128 . [Online]. Available:
[Link]
[105] J. Du, Z. Zhang, and W. Tao, “Multibcd: A multimodal model that simulates the human diagnostic process
for automated breast cancer detection,” Biomedical Signal Processing and Control, vol. 104, p. 107410, Jun.
2025. doi: 10.1016/[Link].2024.107410 . [Online]. Available: [Link]
[106] T. Ekoko, M. S. Ebai, and R. Anang, “P433: Ai-enhanced early detection of breast cancer through
multimodal imaging fusion,” The Breast, vol. 80, p. 104246, Feb. 2025. doi:
10.1016/[Link].2025.104246 . [Online]. Available: [Link]
[107] S. Raichand, V. Blaya-Novakova, S. Berber, A. Livingstone,
N. Noguchi, and N. Houssami, “Digital breast tomosynthesis for breast cancer diagnosis in women with dense
breasts and additional breast cancer risk factors: A systematic review,” The Breast, vol. 77, p. 103767, 2024.
[108] Y. Yu, W. Ren, L. Mao, W. Ouyang, Q. Hu, Q. Yao, Y. Tan, Z. He, X. Ban, H. Hu, R. Lin, Z. Wang, Y.
Chen, Z. Wu, K. Chen, J. Ouyang, T. Li, Z. Zhang, G. Liu, X. Chen, Z. Li, X. Duan, J. Wang, and H. Yao,
“Mri-based multimodal ai model enables prediction of recurrence risk and adjuvant therapy in breast cancer,”
Pharmacological Research, vol. 216, p. 107765, Jun. 2025. doi: 10.1016/[Link].2025.107765. [Online].
Available: [Link]
[109] N. K. Stout, D. L. Miglioretti, Y.-R. Su, C. I. Lee, L. Abraham, O. Alagoz, H. J. de Koning, J. M.
 53

Hampton, L. Henderson, K. P. Lowry et al., “Breast cancer screening using mammography, digital breast
tomosynthesis, and magnetic resonance imaging by breast density,” JAMA Internal Medicine, vol. 184, no. 10,
pp. 1222–1231, 2024.
[110] T. Li, J. Isautier, J. M. Lee, M. L. Marinovich, and N. Houssami, “Per-formance of digital breast tomosynthesis
versus digital mammography in women with a family history of breast cancer: A systematic review,” Clinical
Breast Cancer, vol. 25, no. 2, pp. e103–e112, 2025.
[111] C. O. Lew, M. Harouni, E. R. Kirksey, E. J. Kang, H. Dong, H. Gu, L. J. Grimm, R. Walsh, D. A. Lowell,
and M. A. Mazurowski, “A publicly available deep learning model and dataset for segmentation of breast,
fibroglandular tissue, and vessels in breast mri,” Scientific reports, vol. 14, no. 1, p. 5383, 2024.

[112] E. D. Carvalho, O. P. da Silva Neto, and A. O. de Carvalho Filho, “Deep learning-based tumor segmentation
and classification in breast mri with 3tp method,” Biomedical Signal Processing and Control, vol. 93, p.
106199, 2024.
[113] A. Sultana, “Enhancing breast cancer image analysis through attention mechanisms: A comparative study
of u-net and attention u-net models,” in 2024 IEEE International Conference on Computing, Applications
and Systems (COMPAS). IEEE, 2024, pp. 1–8.
[114] I. Hussain, Z. Lai, H. Kong, and F. Hussain, “Multi-scale feature fusion for breast tumor grading using
dual-adaptive attention mechanisms,” Pattern Recognition, vol. 169, p. 111838, 2026.
[115] C. K. Toa, M. Elsayed, and K. S. Sim, “Deep residual learning with attention mechanism for breast cancer
classification,” Soft Computing, vol. 28, no. 15, pp. 9025–9035, 2024.
[116] A. H. Abdulaal, M. Valizadeh, B. M. Albaker, R. A. Yassin, M. C. Amirani, and A. S. Shah, “Enhancing
breast cancer classification using a modified googlenet architecture with attention mechanism,” Al-Iraqia
Journal for Scientific Engineering Research, vol. 3, no. 1, pp. 47–63, 2024.
[117] S. Anari, S. Sadeghi, G. Sheikhi, R. Ranjbarzadeh, and M. Ben-dechache, “Explainable attention based
breast tumor segmentation using a combination of unet, resnet, densenet, and efficientnet models,”
Scientific Reports, vol. 15, no. 1, p. 1027, 2025.
[118] D. R. Nayak et al., “Rdtnet: A residual deformable attention based transformer network for breast cancer
classification,” Expert Systems with Applications, vol. 249, p. 123569, 2024.
[119] G. Kulandaivelu, A. Taluja, M. Gawas, and R. K. Nath, “Automated breast cancer diagnosis optimized with
higher-order attribute-enhancing heterogeneous graph neural networks using mammogram images,”
Biomedical Signal Processing and Control, vol. 97, p. 106659, 2024.
[120] A. Basaad, S. Basurra, E. Vakaj, M. Aleskandarany, and M. M. Ab-delsamea, “Graphx-net: A graph neural
network-based shapley values for predicting breast cancer occurrence,” IEEE Access, vol. 12, pp. 93
993–94 007, 2024.
[121] E. A. Agyekum, W. Kong, Y.-z. Ren, E. Issaka, J. Baffoe, W. Xian, G. Tan, C. Xiong, Z. Wang, X. Qian et
al., “A comparative analysis of three graph neural network models for predicting axillary lymph node metastasis
in early-stage breast cancer,” Scientific Reports, vol. 15, no. 1, p. 13918, 2025.
[122] Y. Ren, Y. Gao, W. Du, W. Qiao, W. Li, Q. Yang, Y. Liang, and G. Li, “Classifying breast cancer using
multi-view graph neural network based on multi-omics data,” Frontiers in Genetics, vol. 15, p. 1363896,
54

2024.
[123] X. Chang, Z. Zhang, J. Sun, K. Lin, and P. Song, “Breast cancer image classification based on h&e
staining using a causal attention graph neural network model,” Medical & Biological Engineering &
Computing, pp. 1–15, 2025.
[124] N. Saeidi, H. Karshenas, B. Shoushtarian, S. Hatamikia, R. Woitek, and A. Mahbod, “Leveraging medical
foundation model features in graph neural network-based retrieval of breast histopathology images,” arXiv
preprint arXiv:2405.04211, 2024.
[125] S. Lee, J. Park, Y. Piao, D. Lee, D. Lee, and S. Kim, “Multi-layered knowledge graph neural network
reveals pathway-level agreement of three breast cancer multi-gene assays,” Computational and Structural
Biotechnology Journal, vol. 23, pp. 1715–1724, 2024.
[126] A. H. Abdulwahhab, O. Bayat, and A. A. Ibrahim, “Hafmab-net: hierarchical adaptive fusion based on
multilevel attention-enhanced bottleneck neural network for breast histopathological cancer classi-
fication,” Signal, Image and Video Processing, vol. 19, no. 5, p. 410, 2025.
[127] A. Golnari and M. Diba, “Adaptive real-time multi-loss function optimization using dynamic memory
fusion framework: A case study on breast cancer segmentation,” Biomedical Signal Processing and
Control, vol. 111, p. 108388, 2026.
[128] W. Liu and W. Yu, “Deep feature adaptive fusion algorithm for breast cancer histopathology image
classification,” in 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE,
2024,
pp. 5507–5514.
[129] J. Li, K. Wang, and X. Jiang, “Robust multi-subtype identification of breast cancer pathological images
based on a dual-branch frequency domain fusion network,” Sensors, vol. 25, no. 1, p. 240, 2025.
[130] M. A. S. Al Husaini, M. H. Habaebi, and M. R. Islam, “Real-time thermography for breast cancer detection
with deep learning,” Discover Artificial Intelligence, vol. 4, no. 1, p. 57, 2024.
[131] Z. Yang, Y. Ji, Q. Jia, Y. Feng, R. Ji, M. Bai, H. Yan, F. Sun, R. Zhang, and Z. Wang, “Real-time detection
and resection of sentinel lymph node metastasis in breast cancer through a rare earth nanoprobe based nir-
iib fluorescence imaging,” Materials Today Bio, vol. 28, p. 101166, 2024.
 55

VIII. STATEMENTS AND DECLARATIONS/ ETHICS STATEMENTS

Conflict of Interest

No competing interests.

Funding Information

No external funding.

Statement on the use of Generative AI Tools

This manuscript was proofread using ChatGPT-5, an AI-powered language model developed by OpenAI. The tool
was used to refine grammar, enhance clarity, and improve the readability of the text. All technical content, data
interpretations, and conclusions were indepen-dently reviewed and verified by the authors to ensure accuracy and
integrity. The use of ChatGPT-5 was solely for linguistic and stylistic improvements, with no influence on the
scientific or analytical aspects of the manuscript.

Research Involving Human and/or Animals

Not Applicable.

Informed Consent

Not Applicable.

Declaration of interests

☒ The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

☐ The authors declare the following financial interests/personal relationships which may be considered as potential
competing interests:
56

Graphical Abstract