Breast Cancer Detection Using Deep Learning

Perumandla Rohith Goud

Neha Gautam Amit Kumar
University Institue of
University Institue of University Institue of
Engineering Chandigarh
Engineering Chandigarh Engineering Chandigarh
University
University University
Mohali, India
Mohali, India Mohali, India
rohithperumandla9@[Link]
neha.e13217@[Link] amitkumartiwary99@[Link]

Abstract— Breast cancer is a significant global health

In this study, we employ a well-structured CNN
concern, affecting millions of women worldwide. Early and
accurate detection of breast cancer is critical for improving architecture to automatically extract meaningful features from
patient outcomes and reducing mortality rates. Deep learning, breast images. The network is meticulously trained on an
a subset of artificial intelligence, has emerged as a promising extensive and diverse dataset of mammograms, enabling it
tool for enhancing breast cancer detection and diagnosis. This to discern the subtle yet crucial patterns and anomalies
research explores the application of deep learning techniques associated with breast cancer. The trained model is then
to analyze mammograms and other medical imaging data for employed to classify unseen images, providing healthcare
early breast cancer detection. In this study, a convolutional professionals with a rapid and reliable diagnostic tool that can
neural network (CNN) architecture is employed to assist in making informed decisions.
automatically extract meaningful features from breast images.
The findings of this research underscore the immense
The network is trained on a large and diverse dataset of
potential of deep learning in the field of breast cancer
mammograms to learn the subtle patterns and abnormalities
associated with breast cancer. The trained model is then used detection. The CNN model exhibits high accuracy in
to classify new, unseen images as either benign or malignant, categorizing breast cancer cases, offering a ray of hope in
providing a rapid and reliable diagnostic tool for healthcare augmenting the capabilities of radiologists and oncologists for
professionals. The results of this research demonstrate the swifter and more informed clinical decisions. This study
potential of deep learning in breast cancer detection, as the also delves into the challenges and limitations associated with
CNN model achieves high accuracy in classifying breast cancer deep learning-based breast cancer detection, encompassing
cases. The integration of deep learning into clinical practice issues like data scarcity, model interpretability, and ethical
can assist radiologists and oncologists in making more considerations. By addressing these challenges, this
informed and timely decisions, ultimately improving patient research contributes to the ongoing endeavours to harness the
care and survival rates. Furthermore, this study investigates power of deep learning in the early detection and management
the challenges and limitations of deep learning-based breast of breast cancer, ushering in new opportunities for research,
cancer detection, including issues related to data scarcity, development, and clinical implementation.
model interpretability, and ethical considerations. By
addressing these challenges, the research contributes to the II. LITERATURE REVIEW
ongoing efforts to harness the power of deep learning for the
early detection and management of breast cancer, offering new A. Overview of Traditional Breast Cancer
avenues for research, development, and clinical Detection Methods
implementation. Breast cancer detection has traditionally relied on a
combination of clinical examinations, mammography,
Keywords—Convolutional Neural Network (CNN), ultrasound, and biopsy procedures. Mammography, in
Mammograms, Deep Learning Models, Breast Cancer Detection, particular, has been a widely used tool for early breast
DDSM Dataset, Mammographic Image Analysis. cancer detection. This method involves X-ray imaging of
I. INTRODUCTION the breast to identify abnormalities in breast tissue.
However, traditional methods have limitations, including the
Breast cancer is a formidable global health challenge, risk of false positives, discomfort for patients, and exposure
constituting a major health concern for women across the to ionizing radiation..
world. It remains one of the most prevalent malignancies
and a leading cause of cancer-related mortality, B. Previous Research Related to Deep Learning and
emphasizing the paramount importance of timely and Breast Cancer Detection
precise detection. Early diagnosis can significantly Recent years have witnessed a surge in research
influence treatment success and patient survival rates. In this exploring the integration of deep learning techniques for
context, the field of deep learning has emerged as a promising breast cancer detection. Convolutional Neural Networks
and innovative approach to enhancing breast cancer (CNNs) have been at the forefront of these efforts. Various
detection and diagnosis. studies have shown promising results using CNNs to
Deep learning, a subfield of artificial intelligence, has analyse mammograms and other medical images. For
demonstrated remarkable potential in various applications, example, a study by Estevaetal. (2017) demonstrated the
particularly in the field of medical imaging. The ability of efficacy of deep learning models in identifying breast
deep learning algorithms, such as convolutional neural cancer metastasis in lymph nodes. Furthermore, research by
networks (CNNs), to automatically learn and extract Haetal. (2018) applied deep learning to distinguish between
intricate patterns from complex data has opened new benign and malignant breast lesions, achieving high
avenues for improving the accuracy and efficiency of breast accuracy rates.
cancer detection. This research explores the utility of deep Deep learning models, including Recurrent Neural
learning techniques in the analysis of mammograms and Networks (RNNs), have also been used for the analysis of
other medical imaging data, aiming to enhance the early clinical data and electronic health records. These models
detection of breast cancer. can assist in predicting the risk of breast cancer and
tailoring
personalized treatment plans. Research by Leeetal. (2020) cases. F1-score is a harmonic mean of precision and recall,
illustrated the use of RNNs to predict breast cancer survival providing a balanced measure. AUC-ROC quantifies the
outcomes by integrating various patient data, including model's ability to distinguish between benign and malignant
clinical, pathological, and genomic information. cases.
C. Advantages and Limitations of Existing Methods
TABLE I. EVALUATION METRICS FORMULAS
Traditional methods for breast cancer detection, while
established, have limitations. Mammography, although Evaluation Formulas
effective, may miss certain types of breast cancer, especially Metrics
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 + 𝑇𝑟𝑢𝑒
in women with dense breast tissue. It can also produce
𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠
Accuracy
false- positive results, leading to unnecessary anxiety and

𝑇𝑜𝑡𝑎𝑙 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠
additional testing. Moreover, the manual interpretation of
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠
medical images is subject to interobserver variability, which

𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 + 𝐹𝑎𝑙𝑠𝑒

deep learning models aim to mitigate. Precision

𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠
Deep learning methods offer several advantages, including
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠
automation, scalability, and the potential to enhance accuracy.

𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 + 𝐹𝑎𝑙𝑠𝑒

These models can process vast amounts of medical image data Recall

𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠
rapidly, aiding in early detection and reducing the risk of false

2 × 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑅𝑒𝑐𝑎𝑙𝑙
diagnoses. However, they are not without limitations, such
as the need for large labelled datasets, potential biases in F1 Score
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙
training data, and challenges in model interpretability.

III. METHODOLOGY
The success of a deep learning model in breast cancer Additionally, confusion matrices and heatmaps were
detection critically depends on the quality and quantity of generated to gain insights into the model's performance on
the data used. In this study, a diverse dataset of different categories, allowing for the identification of potential
mammographic images was collected from various sources, sources of error.
including medical institutions and publicly available
repositories. The dataset consists of both benign and
malignant cases, ensuring a balanced representation of
different breast conditions. Data preprocessing involved
several essential steps to ensure the data's suitability for
training and testing deep learning models. These steps
included resizing the images to a standardized dimension,
normalizing pixel values to a common scale, and addressing
issues such as noise and artefacts. Furthermore, the dataset
was divided into training, validation, and test sets to
facilitate model training and evaluation. Randomization was
applied to minimize any potential bias in the data partitioning.
This research employed a range of deep-learning models
known for their effectiveness in image classification tasks.
Specifically, convolutional neural networks (CNNs) were
Fig. 1. Confusion Matrix
the primary architecture choice due to their capacity to
capture intricate features within images. Models such as IV. EXPERIMENTS AND RESULTS
VGG-16, Inception, and ResNet were considered for their
deep and well-established architectures. These models were A. Dataset:
adapted to our specific breast cancer detection task by In this study on breast cancer detection using deep
modifying the output layers to correspond to the binary learning, the dataset used is a critical component of the
classification of benign and malignant cases. Transfer research process. It provides the foundational information
learning was employed, using pre-trained weights from necessary for the training and evaluation of deep learning
models trained on large-scale image datasets. This approach models. The dataset utilized in this research was sourced from
allowed the model to leverage the features learned from multiple medical institutions and publicly available
diverse image data and fine-tune these features for breast repositories to ensure diversity and representativeness.
cancer detection. Transfer learning is especially beneficial These sources included renowned medical centres
when working with relatively small medical image datasets. specializing in breast cancer diagnosis, as well as open-
A. Evaluation Metrics access datasets that have been rigorously curated for
research purposes. The data collection process was
The performance of the deep learning models was meticulous to maintain the highest standards of data quality
evaluated using a range of standard metrics to assess their and privacy compliance. The dataset comprises a total of
ability to detect breast cancer accurately. The primary 2620 mammographic images, which were distributed across a
evaluation metrics included accuracy, precision, recall, F1- balanced representation of benign and malignant cases. A
score, and the area under the receiver operating balanced distribution was maintained to prevent class
characteristic curve (AUC-ROC). Accuracy represents the imbalance issues that could affect model training and
overall correctness of the model's predictions, while evaluation. This size was deemed sufficient to train deep
precision and recall measure the model's ability to correctly learning models effectively while keeping computational
classify positive resources and time constraints in mind.
B. Experimental Setup and Configuration: findings of this study indicate the promising diagnostic
The experimental phase of this study was designed to potential of these models. Through the careful utilization of
assess the performance of the deep learning model in breast the DDSM dataset, our deep learning model demonstrated
cancer detection. The study utilized the DDSM (Digital commendable results, as witnessed by metrics such as
Database for Screening Mammography) dataset, which accuracy, precision, recall, F1-score, and ROC AUC.
includes mammographic images. To facilitate
reproducibility and fair comparison, the following The merit of deep learning in this context arises from its
experimental setup and configurations were employed: ability to recognize intricate patterns and features within
medical images, resulting in heightened diagnostic
Dataset: The DDSM dataset contains a diverse accuracy. The comparison with traditional detection
collection of mammographic images, including both mass methods accentuated the model's superiority, offering a
and calcification cases. It was divided into a training set and glimpse of the efficiency, automation, and improved
a test set, with a respective 80% and 20% split, maintaining diagnostic capabilities that deep learning could usher into
class balance to avoid data bias. the field of breast cancer detection.
Data Preprocessing: Prior to model training, extensive data However, it is imperative to acknowledge that the
preprocessing was performed. This included image resizing to translation of research findings into clinical practice
a consistent dimension (e.g., 224x224), normalization of pixel necessitates rigorous validation, adherence to medical
values, and augmentation techniques, such as random regulations, and ethical considerations. The transition to
rotations and flips. These preprocessing steps aim to ensure real- world deployment requires meticulous examination,
data uniformity and enhance model robustness. including access to larger and more diverse datasets,
rigorous model validation, and collaboration with medical
Deep Learning Model: A convolutional neural network
practitioners to ensure the highest standards of patient care.
(CNN) architecture was chosen for its effectiveness in
image classification tasks. The model architecture was REFERENCES
based on a pre-trained network, fine-tuned for binary
[1] F. Azour and A. Boukerche, "Design Guidelines for Mammogram-
classification of benign and malignant lesions. The chosen Based Computer-Aided Systems Using Deep Learning Techniques,"
optimizer was Adam, and the model was trained with a in IEEE Access, vol. 10, pp. 21701-21726, 2022, doi:
batch size of 32 for a predefined number of epochs. 10.1109/ACCESS.2022.3151830.
[2] Suckling, J., Parker, J., Dance, D. et al. (2015). Mammographic Image
Evaluation Metrics: The primary evaluation metrics Analysis Society (MIAS) database v1.21. [Dataset]. Apollo -
used to gauge the model's performance were accuracy, University of Cambridge Repository.
precision, recall, F1-score, and the area under the receiver [Link]
operating characteristic curve (AUC-ROC). These metrics [3] Z. Wang et al., "Breast Cancer Detection Using Extreme Learning
provide a comprehensive assessment of the model's Machine Based on Feature Fusion With CNN Deep Features," in
diagnostic capabilities. IEEE Access, vol. 7, pp. 105146-105158, 2019, doi:
10.1109/ACCESS.2019.2892795.
C. Presentation of Results: [4] I. Hirra et al., "Breast Cancer Classification From Histopathological
Images Using Patch-Based Deep Learning Modeling," in IEEE
The model's performance was evaluated on the test set Access, vol. 9, pp. 24273-24287, 2021, doi:
to determine its ability to accurately classify breast lesions. 10.1109/ACCESS.2021.3056516.
The results are presented as follows: [5] D. Tsietso et al., "Multi-Input Deep Learning Approach for Breast
Cancer Screening Using Thermal Infrared Imaging and Clinical
Accuracy: The accuracy of the model on the test set was Data," in IEEE Access, vol. 11, pp. 52101-52116, 2023, doi:
0.5350, indicating the proportion of correct predictions out 10.1109/ACCESS.2023.3280422..
of the total samples. [6] S. Sharmin, T. Ahammad, M. A. Talukder and P. Ghose, "A Hybrid
Dependable Deep Feature Extraction and Ensemble-Based Machine
Precision: The precision score, measuring 0.5441, Learning Approach for Breast Cancer Detection," in IEEE Access,
reflects the model's ability to correctly identify malignant vol. 11, pp. 87694-87708, 2023, doi:
cases while minimizing false positives. 10.1109/ACCESS.2023.3304628.
[7] N. Wu et al., "Deep Neural Networks Improve Radiologists’
Recall: The recall score, 0.3737, represents the model's Performance in Breast Cancer Screening," in IEEE Transactions on
capability to identify most of the actual malignant cases Medical Imaging, vol. 39, no. 4, pp. 1184-1194, April 2020, doi:
without missing many, reducing false negatives. 10.1109/TMI.2019.2945514.
[8] V. Patel, V. Chaurasia, R. Mahadeva and S. P. Patole, "GARL-Net:
F1-Score: The F1-score, which combines precision and Graph Based Adaptive Regularized Learning Deep Network for
recall into a single metric, was 0.4431. It highlights the Breast Cancer Classification," in IEEE Access, vol. 11, pp. 9095-
model's overall performance in the context of a binary 9112, 2023, doi: 10.1109/ACCESS.2023.3239671.
classification task. [9] J. Zheng, D. Lin, Z. Gao, S. Wang, M. He and J. Fan, "Deep Learning
Assisted Efficient AdaBoost Algorithm for Breast Cancer Detection
ROC Curves: The model's receiver operating and Early Diagnosis," in IEEE Access, vol. 8, pp. 96946-96954, 2020,
characteristic (ROC) curve visually portrays its ability to doi: 10.1109/ACCESS.2020.2993536.
distinguish between benign and malignant cases. The AUC- [10] G. Li, C. Li, G. Wu, D. Ji and H. Zhang, "Multi-View Attention-
ROC score, 0.5378, quantifies this ability, with a higher Guided Multiple Instance Detection Network for Interpretable Breast
value indicating better discrimination. Cancer Histopathological Image Diagnosis," in IEEE Access, vol. 9,
pp. 79671-79684, 2021, doi: 10.1109/ACCESS.2021.3084360.
CONCLUSION [11] J. Xu et al., "Stacked Sparse Autoencoder (SSAE) for Nuclei
Detection on Breast Cancer Histopathology Images," in IEEE
In the pursuit of enhancing breast cancer detection, this Transactions on Medical Imaging, vol. 35, no. 1, pp. 119-130, Jan.
research delved into the application of deep learning models 2016, doi: 10.1109/TMI.2015.2458702.
within the realm of mammographic image analysis. The [12] S. Rajpal, M. Agarwal, V. Kumar, A. Gupta and N. Kumar, "Triphasic
DeepBRCA-A Deep Learning-Based Framework for Identification of
Biomarkers for Breast Cancer Stratification," in IEEE Access, vol. 9,
pp. 103347-103364, 2021, doi: 10.1109/ACCESS.2021.3093616.
[13] A. Saber, M. Sakr, O. M. Abo-Seida, A. Keshk and H. Chen, "A Nuclei in Breast Cancer Evaluation," in IEEE Transactions on Image
Novel Deep-Learning Model for Automatic Detection and Processing, vol. 27, no. 5, pp. 2189-2200, May 2018, doi:
Classification of Breast Cancer Using the Transfer-Learning 10.1109/TIP.2018.2795742.
Technique," in IEEE Access, vol. 9, pp. 71194-71209, 2021, doi:
10.1109/ACCESS.2021.3079204. [17] S. Misra et al., "Bi-Modal Transfer Learning for Classifying Breast
Cancers via Combined B-Mode and Ultrasound Strain Imaging," in
[14] M. Sebai, T. Wang and S. A. Al-Fadhli, "PartMitosis: A Partially IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency
Supervised Deep Learning Framework for Mitosis Detection in Breast Control, vol. 69, no. 1, pp. 222-232, Jan. 2022, doi:
Cancer Histopathology Images," in IEEE Access, vol. 8, pp. 45133- 10.1109/TUFFC.2021.3119251.
45147, 2020, doi: 10.1109/ACCESS.2020.2978754.
[18] R. Singh, T. Ahmed, A. Kumar, A. K. Singh, A. K. Pandey and S. K.
[15] E. Strelcenia and S. Prakoonwit, "Improving Cancer Detection Singh, "Imbalanced Breast Cancer Classification Using Transfer
Classification Performance Using GANs in Breast Cancer Data," in Learning," in IEEE/ACM Transactions on Computational Biology and
IEEE Access, vol. 11, pp. 71594-71615, 2023, doi: Bioinformatics, vol. 18, no. 1, pp. 83-93, 1 Jan.-Feb. 2021, doi:
10.1109/ACCESS.2023.3291336. 10.1109/TCBB.2020.2980831.
[16] M. Saha and C. Chakraborty, "Her2Net: A Deep Framework for
Semantic Segmentation and Classification of Cell Membranes and