CAN LARGE LANGUAGE MODELS CHALLENGE CNNS IN MEDICAL IMAGE ANALYSIS?

Shibbir Ahmed Shahnewaz Karim Sakib Anindya Bijoy Das

Texas State University University of Tennessee at Chattanooga The University of Akron
San Marcos, TX, USA Chattanooga, TN, USA Akron, OH, USA

ABSTRACT In parallel with these imaging advances, large language

models (LLMs) [8] have emerged as powerful tools for pro-
This study presents a multimodal AI framework designed for cessing and interpreting clinical narratives, radiology reports,
arXiv:2505.23503v2 [[Link]] 3 Jun 2025

precisely classifying medical diagnostic images. Utilizing and other unstructured data sources [8, 9, 10, 11]. LLMs com-
publicly available datasets, the proposed system compares the plement image-based predictions to boost diagnostic accuracy
strengths of convolutional neural networks (CNNs) and differ- by extracting and contextualizing patient information. For
ent large language models (LLMs). This in-depth comparative example, integrating LLMs with imaging architectures enables
analysis highlights key differences in diagnostic performance, automated report generation and decision support, i.e., cross-
execution efficiency, and environmental impacts. Model evalu- referencing clinical findings with current literature, to reduce
ation was based on accuracy, F1-score, average execution time, clinician workload and diagnostic errors [9]. Furthermore,
average energy consumption, and estimated CO2 emission. the synergy between LLMs and advanced machine learning
The findings indicate that although CNN-based models can techniques enhances AI interpretability and explainability [12],
outperform various multimodal techniques that incorporate improving trust among healthcare professionals and enabling
both images and contextual information, applying additional personalized treatments. Incorporating LLMs augments imag-
filtering on top of LLMs can lead to substantial performance ing model performance, streamlines diagnosis, and improves
gains. These findings highlight the transformative potential of clinical workflows, ultimately enhancing patient care.
multimodal AI systems to enhance the reliability, efficiency, In this study, we investigate the reliability and performance
and scalability of medical diagnostics in clinical settings. of multimodal AI approaches for diagnostic image classifica-
tion by comparing the effectiveness of CNN-based architec-
Index Terms— Medical Diagnostics, Multimodal AI, Con- tures with that of several benchmark large language models
volutional Neural Networks, Large Language Models, Medical [13, 14], including OpenAI’s GPT-4o, and Meta’s LLaMA. We
Image Classification employ publicly available datasets comprising X-ray, CT, and
MRI modalities (as shown in Fig. 1), which are meticulously
1. INTRODUCTION partitioned into training, validation, and unseen testing sets. In
The rapid evolution of artificial intelligence (AI) has ushered addition, we explore these multiple LLMs to assess their poten-
in transformative approaches for medical diagnostics, particu- tial in diagnostics, where few studies [15] have concurrently
larly in the analysis of diagnostic imaging data [1]. Over the addressed both utility and reliability.
past decade, machine learning techniques have been applied Our systematic evaluation employs performance metrics
successfully to tasks such as image classification, segmenta- such as accuracy, F1-score, average execution time, average
tion, and anomaly detection in modalities including X-ray [2], energy consumption, and estimated CO2 emission to provide
computed tomography (CT) [3], and magnetic resonance imag- detailed insights into the strengths and limitations of CNNs
ing (MRI) [4]. Early breakthroughs using convolutional neural and LLMs in real-world clinical scenarios. Additionally, we
networks (CNNs) [2], recurrent neural networks (RNNs), and explore confidence scores as a measure of reliability, offering
autoencoders demonstrated the potential of deep architectures a comparative analysis of the trustworthiness of these models
to learn hierarchical representations directly from raw images. in diagnostic decision-making. This study contributes by:
Recent advancements in deep learning have introduced alter- • Benchmarking CNNs and LLMs on classifying diverse
native architectures such as transformers [5], capsule networks medical imaging datasets, including image analysis by
[6], and diffusion models [7], significantly enhancing the ca- the LLM models for future study.
pabilities of medical image analysis. These innovative models
excel in capturing complex patterns and dependencies in med- • Analyzing computational efficiency, including runtime,
ical data, which is crucial for accurate diagnosis and effective energy consumption, and associated CO2 emissions.
treatment planning. Their successful application across various • Highlighting trade-offs between model accuracy and
clinical scenarios demonstrates their potential in improving resource demands, emphasizing considerations critical
the precision and efficiency of healthcare interventions. for practical and sustainable AI adoption in healthcare.
 (a) Chest X-ray (b) Chest CT scan (c) Brain MRI

Fig. 1. Different Medical Imaging Modalities and Pathologies: (a) Chest X-rays depicting normal lung anatomy, COVID-19 infection, lung
opacity, and viral pneumonia. (b) Chest CT scans for normal lung, squamous cell carcinoma, adenocarcinoma, and large cell carcinoma.
(c) Brain MRI scans highlighting conditions such as no tumor, glioma, meningioma, and pituitary abnormalities. These exemplify typical
radiological features used for training and evaluating AI-driven diagnostic models in medical imaging.

In summary, our research findings provide useful bench- 2.2. Related Works
marks for researchers and valuable insights for medical pro-
fessionals seeking to adopt AI-driven diagnosis, ultimately Next, we explore relevant studies focused on the analysis and
improving clinical workflow reliability and efficiency. classification of medical images, such as X-rays, CT scans,
and MRI data, utilizing various LLM-based methodologies.
2. DATASETS AND RELATED WORKS Transfer Learning and Pre-trained Models: Recent studies
have adapted pre-trained language models, like BERT and
In this section, we detail the datasets utilized in this study GPT, to extract and contextualize features from medical im-
and review existing literature relevant to the classification of ages. For example, [20] developed MMBERT explores on
medical images. two visual question answering (VQA) datasets for radiology
images, outperforms different previous ensemble models and
2.1. Dataset Selection provides attention maps which help in model interpretability.
To ensure a comprehensive evaluation, we select the follow- Multi-modal Learning Approaches: Multi-modal learning
ing publicly available datasets covering a range of imaging techniques have shown great promise in enhancing the analysis
modalities and diagnostic tasks. These datasets were curated of medical images by integrating models trained on different
to maintain class balance, diversity, and representativeness. types of data [21]. For instance, incorporating neural networks
Chest X-Ray Dataset: The COVID-19 Radiography Database with LLMs can enable a synergistic approach where the model
[16] contains X-ray images labeled as COVID-19, pneumonia, benefits from the spatial recognition capabilities of neural
lung opacity, and normal. This dataset is commonly used to networks while leveraging the contextual processing strengths
benchmark AI models for classifying chest X-rays. [17]. of LLMs. The work in [22] demonstrates the effectiveness of
large visual language models in analyzing biomedical images
Brain Tumor MRI Dataset: The Brain Tumor MRI Dataset such as brain MRIs, microscopic images of blood cells, and
[18], available on Kaggle, includes T1-weighted MRI images chest X-rays.
labeled as glioma, meningioma, pituitary tumor, or no tumor.
This dataset enables the evaluation of models in brain tumor Fine-tuning on the Datasets: Tailoring LLMs to specific
classification tasks. medical tasks has proven to be an effective strategy [22]. Fine-
tuning models on specialized datasets, such as radiology re-
Chest CT Scan Dataset: This dataset [19], also available on
ports, enhances their ability to accurately classify medical
Kaggle, includes chest CT scans labeled as normal, adenocar-
images, including CT scans of lung nodules.
cinoma, large cell carcinoma and squamous cell carcinoma.
This enables us to work on different chest cancer detection. Explainability and Ethical Considerations: With AI’s grow-
Sample images are illustrated in Fig. 1, which demon- ing role in healthcare, the demand for explainable AI models
strates the distinctions between different types of medical has surged [23]. Explanations generated by LLMs can help
imaging: chest X-ray in Fig. 1(a), chest CT scan in Fig. 1(b) radiologists understand the AI’s reasoning process, potentially
and brain MRI in Fig. 1(c). increasing trust and reliability in AI-assisted diagnostics [12].
 3. ANALYSIS OF THE MODELS 1. Initial Label-Based Filtering: We begin by enumerat-
ing the responses across the entire training set. From this
In this section, we present the various models employed in this dataset, we first extract only those samples correspond-
study for classifying medical images, followed by an analysis ing to the desired label (e.g., “normal” in the context of
of the corresponding classification processes. COVID classification).
3.1. Classification Approaches 2. Confidence Thresholding: To further refine the dataset,
we associate each sample with a confidence score (such
In this study, we investigate two complementary paradigms for as the output of a softmax layer reflecting the model’s
medical image classification: large language models (LLMs) certainty in its prediction). We then discard samples
and deep neural networks (DNNs). Each approach is tailored with scores below a chosen threshold (e.g., 0.8), thereby
to address the unique challenges of medical imaging and is retaining only high-confidence, label-consistent samples
described in detail below. that are more likely to be correctly classified and less
Large Language Models (LLMs): Traditionally employed prone to noise or ambiguity.
for natural language processing tasks, LLMs have recently
3. Context and Feature Extraction: Next, we examine
shown promise in multimodal applications, including image
these high-quality samples to identify key contextual
classification. By leveraging language-based embeddings, im-
elements and features (e.g., specific words, phrases,
ages can be transformed into descriptive representations that
or structural patterns) that enable the Large Language
LLMs process to generate classification predictions. In this
Model (LLM) to classify them as the chosen label.
study, we explore methods to integrate image features directly
into LLM architectures, focusing on adapting state-of-the- 4. Context Aggregation and Question Formulation: We
art models such as GPT-4o and Llama3.2-vision for medi- aggregate these contextual insights from all filtered sam-
cal diagnostics. We assess LLM feasibility in tasks typically ples and present them to GPT-4o. This step produces a
dominated by image-centric models, identifying strengths and consolidated understanding of the features driving clas-
limitations in handling visual medical data. sification decisions, as well as a set of targeted questions
Deep Neural Networks (DNNs): DNNs, particularly convo- whose answers are critical for label determination.
lutional neural networks (CNNs), have established themselves 5. Application During Testing: Finally, in the testing
as a gold standard for image-based tasks, including medical phase, we incorporate these targeted questions into the
diagnostics. For this study, we use state-of-the-art CNN archi- user prompt, rather than attempting to enumerate all
tectures that have demonstrated high performance in medical possible features. By prompting the LLM to respond to
imaging, such as the COVID-Roentgen CNN for chest X-rays the key questions directly, we streamline the classifica-
and a TensorFlow-based CNN for brain tumor classification. tion process and increase the reliability of determining
Specifically, for the chest X-Ray dataset, we employed whether a given image matches the target label.
the COVID CNN model and optimized to classify COVID-19 A sample user prompt is provided in Fig. 2. Note that we
and normal chest X-rays. Next, for brain tumor MRI classifi- also include a confidence score for each test image, thereby
cation, we used a TensorFlow-based CNN model capable of quantifying the model’s reliability in its predictions.
distinguishing benign (normal) and three different types of ma-
lignant (cancerous) growths in high-resolution images. For the Please evaluate the chest X-ray image using the following
chest CT scans, we utilized a fine-tuned CNN trained on the radiological criteria:
dataset to differentiate normal and three different chest cancer 1. Clear Lung Fields
scans. All these DNNs are fine-tuned on the respective datasets 2. Normal Heart and Mediastinum
3. Healthy Diaphragm
to ensure optimal performance, emphasizing their ability to 4. Intact Bony Structures
learn domain-specific features from relatively small datasets. 5. Empty Pleural Spaces
Note that each dataset was split into training (80%), valida- Based on the above criteria, does this image appear normal
or COVID-affected?
tion (20%), and unknown (20%) subsets, following standard Please reply in exactly two lines. The first line should
practices for medical image analysis. contain only either normal or abnormal. The second line
should be in the format Confidence Score: <number>
where <number> is a decimal value between 0 and 1.
3.2. Enhanced Data Filtering for LLM Classifications
Now we refine the data filtering process for LLM classification Fig. 2. Context Aggregation and Question Formulation
by implementing a multi-stage strategy that carefully screens
and integrates contextual and feature details. This approach This method streamlines the decision-making process by
ensures only essential data (or features) is selected for analysis, equipping the Large Language Model (LLM) with specific
thereby improving both the precision and consistency of clas- contextual information necessary for the classification task.
sification outcomes. The enhanced methodology is outlined Consequently, this targeted data provision not only acceler-
through the following step-by-step workflow: ates the execution time but also lowers the energy required,
enhancing the model’s efficiency and environmental footprint.
Table 1. Performance comparison of LLM and CNN models in Table 2. Comparison of performance with and without data filtering.
terms of accuracy, F-1 score and average confidence score (CS) across The symbol ↑ indicates a desirable increase in the metric, whereas ↓
different medical imaging modalities indicates a desirable decrease.

Dataset Model Acc. F-1 Avg. CS Metric w/o filtering with filtering
CNN 0.83 0.83 0.79 Accuracy 62% 82.01% ↑
Chest X-ray GPT-4o 0.62 0.54 0.93 Avg. Confid. Score 0.93 0.93 =
Llama3.2-vision 0.65 0.64 0.92 Avg. Execut. Time 6.23 s 2.35 s ↓
CNN 0.98 0.99 0.99 Avg. Energy Consump. 1.84 W-H 1.65 W-H ↓
MRI GPT-4o 0.60 0.58 0.93
Llama3.2-vision 0.52 0.51 0.85
CNN 0.91 0.90 0.94 confidence score (0.93), GPT-4o exhibits a mismatch between
Chest CT GPT-4o 0.22 0.14 0.91 confidence and actual performance.
Llama3.2-vision 0.50 0.48 0.80 Chest CT Scan: For chest CT scans, CNNs again lead in
performance, achieving an accuracy and F1-score of around
4. RESULTS AND DISCUSSIONS 0.91. GPT-4o performs poorly with an accuracy of 0.22 and
an F1-score of 0.14, despite maintaining a confidence score
4.1. Evaluation Metrics
of 0.91. Llama3.2-vision performs slightly better, with an
The performance of CNNs and LLMs was evaluated using sev- accuracy of 0.50 and an F1-score of 0.48, but both LLMs lag
eral key metrics. Accuracy is measured by the percentage of significantly behind CNNs for this modality.
correctly classified samples, providing an overall performance
indicator. Precision assessed the proportion of true positives 4.3. Computational Efficiency Analysis
among predicted positives, minimizing false positives, while The Fig. 3 compares CNNs, GPT-4o, and Llama3.2-vision
Recall measured the proportion of true positives among actual across three datasets (Chest X-ray, MRI, and Chest CT scan)
positives, emphasizing the model’s ability to identify relevant in terms of execution time, energy consumption, and CO2
cases. The F1-Score [24], as the harmonic mean of preci- emissions. We calculated energy consumption and CO2 emis-
sion and recall, offered a balanced measure of performance. sions [26] from execution time, power usage, and average
Additionally, the Confidence Score [25] analyzed prediction carbon intensity. CNNs demonstrate the best computational
reliability using calibration curves. Finally, resource consump- efficiency, with minimal execution time, energy usage, and
tion, such as inference time and energy usage, was evaluated environmental impact, making them highly suitable for real-
comprehensively to assess the models’ efficiency and practi- time medical diagnostics. In contrast, LLMs (GPT-4o and
cality in real-world healthcare applications. Llama3.2-vision) exhibit significantly higher execution times
and energy consumption, particularly Llama3.2-vision, which
4.2. Results Overview
incurs higher costs across all metrics. CNN inference times for
Table 1 summarizes comparative performance of CNNs and chest CT scans can be longer due to complexity of volumetric
LLMs (GPT-4o and Llama3.2-vision) across medical imaging images. These results illustrate the trade-offs between CNNs’
datasets: chest X-ray, Brain MRI, and chest CT scans. The computational efficiency and LLMs’ multimodal reasoning, in-
results indicate notable variations in accuracy, F1-score, and dicating that CNNs are better for scalable clinical applications,
confidence, revealing each model’s strengths and limitations. while LLMs excel in complex contextual tasks.
Chest X-ray: The CNN achieves the highest overall perfor- 4.4. Effect of Enhanced Data Filtering
mance, with precision, recall, accuracy, and F1-score of 0.83,
surpassing both GPT-4o and Llama3.2-vision. Its average con- The results of our enhanced data filtering process, shown in
fidence score of 0.79 indicates reliable prediction confidence. Table 2, highlight its effectiveness with the chest X-ray dataset.
GPT-4o, while achieving the highest confidence score (0.93), Using GPT-4o alone yields 62% accuracy, but adding a filter-
has a relatively low accuracy (0.62) and F1-score (0.54), re- ing step (Sec. 3.2) boosts accuracy to 82% with an average
flecting overconfidence in its predictions. Llama3.2-vision confidence score of 0.93. This filtering also cuts execution
shows slightly better accuracy (0.65) and F1-score (0.64) than time from 6.23 seconds to 2.35 seconds and reduces energy
GPT-4o, but it still underperforms compared to the CNN. consumption from 1.84 watt-hours to 1.65 watt-hours. We an-
Brain MRI: CNNs demonstrate exceptional performance on ticipate similar gains with other LLMs across various imaging
MRI datasets, with near-perfect precision, recall, accuracy, and modalities, but results may vary by dataset complexity.
F1-score (all above 0.98). This underscores their capability 4.5. Additional Insights
to handle well-defined imaging patterns. In contrast, GPT-
4o and Llama3.2-vision exhibit poor accuracy and F1-scores Limitations of traditional LLMs: While the LLMs, includ-
(below 0.60), suggesting limited effectiveness. Despite a high ing GPT-4o and Llama3.2-vision, show moderate success in
 (a) Average Execution Time (b) Average Energy Consumption (c) Average CO2 Emission

Fig. 3. Comparative performance of CNNs and LLMs across three medical imaging datasets: chest X-ray, MRI, and chest CT scans

handling medical image classification tasks, their confidence in medical imaging. Furthermore, explainable AI techniques
scores, particularly for GPT-4o, are disproportionately high rel- can enhance transparency and trust among clinicians, aiding
ative to accuracy, highlighting need for improved calibration. the real-world adoption of these technologies in healthcare.
The lower performance of LLMs suggests that they require
further adaptation or fine-tuning for image-specific tasks. 6. REFERENCES
Strengths of Data Filtering in LLM Providing necessary
[1] S. Kevin Zhou, Hayit Greenspan, Christos Davatzikos,
contextual details to the Large Language Model (LLM), a
James S. Duncan, Bram Van Ginneken, Anant Madab-
framework can simplify the decision-making cycle, reduce
hushi, et al., “A review of deep learning in medical imag-
processing times and energy usage, thereby boosting overall
ing: Imaging traits, technology trends, case studies with
efficiency and lower environmental impact. Such a streamlined
progress highlights, and future promises,” Proceedings
approach fosters more robust and eco-friendly medical image
of the IEEE, vol. 109, no. 5, pp. 820–838, 2021.
classification while maintaining high accuracy.
Insights on Model Calibration: The calibration of confi- [2] Yun Yang, Yuanyuan Hu, Xingyi Zhang, and Song Wang,
dence scores is a critical aspect of deploying AI systems in “Two-stage selective ensemble of CNN via deep tree train-
clinical settings. CNNs demonstrate better alignment between ing for medical image classification,” IEEE Transactions
confidence and accuracy, making them more suitable for high- on Cybernetics, vol. 52, no. 9, pp. 9194–9207, 2022.
stakes applications. LLMs need calibration to enhance trust in [3] Rajesh Kumar, Abdullah Aman Khan, Jay Kumar, Za-
clinical workflows. kria, Noorbakhsh Amiri Golilarz, Simin Zhang, Yang
Dataset Complexity: The performance differences across Ting, Chengyu Zheng, and Wenyong Wang, “Blockchain-
datasets reveal that MRI datasets achieve higher accuracy due federated-learning and deep learning models for COVID-
to distinct and consistent patterns, while chest CT scans are 19 detection using CT imaging,” IEEE Sensors Journal,
more challenging because of complex and subtle features, vol. 21, no. 14, pp. 16301–16314, 2021.
highlighting each model’s strengths and limitations.
[4] Mohammad Ashraf Ottom, Hanif Abdul Rahman, and
Ivo D. Dinov, “Znet: Deep learning approach for 2D
5. CONCLUSION AND FUTURE WORK MRI brain tumor segmentation,” IEEE Jour. Transl. Eng.
Health Med, vol. 10, pp. 1–8, 2022.
This study compared CNNs and LLMs for medical image clas-
sification, highlighting their strengths and limitations. CNNs [5] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong
demonstrated superior accuracy and efficiency, making them Yang, Andriy Myronenko, Bennett Landman, Holger R
well-suited for real-time clinical applications, while LLMs Roth, and Daguang Xu, “UNETR: Transformers for
showed potential in multimodal reasoning but suffered from 3D medical image segmentation,” in Proceedings of the
high computational costs. Our findings emphasize the trade- IEEE/CVF winter conference on applications of com-
offs between accuracy, reliability, and resource consumption, puter vision, 2022, pp. 574–584.
offering practical insights for AI deployment in healthcare. [6] A. F. M. Saif, Tamjid Imtiaz, Shahriar Rifat, Celia Shah-
Future works will focus on hybrid models that merge deep naz, Wei-Ping Zhu, and M. Omair Ahmad, “CapsCovNet:
neural networks with LLMs, enhancing medical diagnostics A modified capsule network to diagnose COVID-19 from
through multimodal AI [27]. This integration aims to improve multimodal medical imaging,” IEEE Transactions on
both accuracy and interpretability by combining visual fea- Artificial Intelligence, vol. 2, no. 6, pp. 608–617, 2021.
tures with contextual reasoning. Additionally, our research
offers a benchmark with LLM-predicted labels and detailed [7] Muzaffer Özbey, Onat Dalmaz, Salman U. H. Dar,
medical image analysis, paving the way for new AI research Hasan A. Bedel, Şaban Özturk, Alper Güngör, and Tolga
 Çukur, “Unsupervised medical image translation with [17] Tawsifur Rahman, Amith Khandakar, Yazan Qiblawey,
adversarial diffusion models,” IEEE Transactions on Anas Tahir, Serkan Kiranyaz, et al., “Exploring the effect
Medical Imaging, vol. 42, no. 12, pp. 3524–3539, 2023. of image enhancement techniques on COVID-19 detec-
tion using chest X-ray images,” Computers in biology
[8] Arun James Thirunavukarasu, Darren Shu Jeng Ting, and medicine, vol. 132, pp. 104319, 2021.
Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan,
and Daniel Shu Wei Ting, “Large language models in [18] Masoud Nickparvar, “Brain tumor MRI dataset,” Kaggle,
medicine,” Nature medicine, vol. 29, no. 8, pp. 1930– 2020, [Link]
1940, 2023. nickparvar/brain-tumor-mri-dataset.

[9] Paul Hager, Friederike Jungmann, Robbie Holland, Ku- [19] Mohamed Hany, “Chest CT-scan images dataset,” Kag-
nal Bhagat, Inga Hubrecht, Manuel Knauer, et al., “Eval- gle, 2021, [Link]
uation and mitigation of the limitations of large language medhanyyy/chest-ctscan-images.
models in clinical decision-making,” Nature medicine,
[20] Yash Khare, Viraj Bagal, Minesh Mathew, Adithi Devi,
vol. 30, pp. 2613–2622, 2024.
U Deva Priyakumar, and CV Jawahar, “MMBERT: Mul-
[10] Anindya Bijoy Das, Shibbir Ahmed, and Shah- timodal BERT pretraining for improved medical VQA,”
newaz Karim Sakib, “Hallucinations and key information in 18th International Symposium on Biomedical Imaging
extraction in medical texts: A comprehensive assess- (ISBI). IEEE, 2021, pp. 1033–1036.
ment of open-source large language models,” preprint,
[21] Ke Zhang, Yan Yang, Jun Yu, Hanliang Jiang, Jianping
arXiv:2504.19061, 2025.
Fan, Qingming Huang, and Weidong Han, “Multi-task
[11] Tiffany H Kung, Morgan Cheatham, Arielle Medenilla, paired masking with alignment modeling for medical
Czarina Sillos, et al., “Performance of ChatGPT on vision-language pre-training,” IEEE Transactions on
USMLE: potential for AI-assisted medical education us- Multimedia, vol. 26, pp. 4706–4721, 2024.
ing large language models,” PLoS digital health, 2023. [22] Minh–Hao Van, Prateek Verma, and Xintao Wu, “On
[12] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, large visual language models for medical imaging analy-
Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, sis: An empirical study,” in IEEE/ACM Conference on
and Mengnan Du, “Explainability for large language Connected Health: Applications, Systems and Engineer-
models: A survey,” ACM Transactions on Intelligent ing Technologies (CHASE), 2024, pp. 172–176.
Systems and Technology, vol. 15, no. 2, pp. 1–38, 2024. [23] Zabir Al Nazi and Wei Peng, “Large language mod-
els in healthcare and medical domain: A review,” in
[13] Seowoo Lee, Jiwon Youn, Hyungjin Kim, Mansu Kim,
Informatics. MDPI, 2024, vol. 11, p. 57.
and Soon Ho Yoon, “CXR-LLaVA: a multimodal large
language model for interpreting chest X-ray images,” Eu- [24] Shahnewaz Karim Sakib and Anindya Bijoy Das, “Ex-
ropean Radiology, pp. 1–13, 2025. plainable vertical federated learning for healthcare: En-
suring privacy and optimal accuracy,” in IEEE Inter-
[14] Dongyuan Wu, Liming Nie, Rao Asad Mumtaz, and
national Conference on Big Data (BigData), 2024, pp.
Kadambri Agarwal, “A LLM-based hybrid-transformer
5068–5077.
diagnosis system in healthcare,” IEEE Journal of Biomed-
ical and Health Informatics, pp. 1–12, 2024. [25] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Wein-
berger, “On calibration of modern neural networks,” in
[15] Zihao Zhao, Sheng Wang, Jinchen Gu, Yitao Zhu, International Conference on Machine Learning (ICML),
Lanzhuju Mei, Zixu Zhuang, Zhiming Cui, et al., “Chat- 2017, pp. 1321–1330.
CAD+: Toward a universal and reliable interactive CAD
using LLMs,” IEEE Transactions on Medical Imaging, [26] Alexandra Sasha Luccioni, Sylvain Viguier, and Anne-
vol. 43, no. 11, pp. 3755–3766, 2024. Laure Ligozat, “Estimating the carbon footprint of
bloom, a 176b parameter language model,” J. Mach.
[16] Muhammad EH Chowdhury, Tawsifur Rahman, Amith Learn. Res., vol. 24, no. 1, Jan. 2023.
Khandakar, Rashid Mazhar, Muhammad Abdul Kadir,
Zaid Bin Mahbub, Khandakar Reajul Islam, Muham- [27] Ekin Tiu, Ellie Talius, Pujan Patel, et al., “Expert-level
mad Salman Khan, Atif Iqbal, Nasser Al Emadi, et al., detection of pathologies from unannotated chest x-ray
“Can AI help in screening viral and COVID-19 pneumo- images via self-supervised learning,” Nature Biomedical
nia?,” IEEE Access, vol. 8, pp. 132665–132676, 2020. Engineering, vol. 6, pp. 1399–1406, 2022.