Pith. sign in

REVIEW 5 major objections 6 minor 41 references

Can Large Language Models Challenge CNNs in Medical Image Analysis?

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CNNs still beat LLMs on medical image classification, but a prompt-filtering step lifts GPT-4o's chest X-ray accuracy from 62% to 82%.

desk verdict The filtering claim is unsupported as reported: the with-filtering prompt is binary while the baseline is undefined, so the 20-point gain may just be task simplification. read the letter →

arxiv 2505.23503 v2 pith:7B4ZRODO submitted 2025-05-29 eess.IV cs.AIcs.CV

Shibbir Ahmed , Shahnewaz Karim Sakib , Anindya Bijoy Das This is my paper
classification eess.IVcs.AIcs.CV
keywords medicalimageclassificationconvolutionalneuralnetworkslargelanguagemodelsmultimodalAIchestX-raybrainMRICTpromptfiltering
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether large language models can compete with convolutional neural networks at classifying medical images, and answers with a qualified no. Across chest X-ray, brain MRI, and chest CT datasets, fine-tuned CNNs reach higher accuracy, F1, and confidence calibration than GPT-4o and Llama3.2-vision, while using far less time and energy. The paper's main positive result is that a multi-stage filtering pipeline—selecting high-confidence label-consistent training samples, extracting their contextual features, and turning them into targeted questions for the LLM—raises GPT-4o's chest X-ray accuracy from 62% to 82% and cuts execution time by roughly two-thirds. That result matters because it suggests a cheap inference-time intervention can substantially improve LLM diagnostics without retraining.

What carries the argument

The central mechanism is the enhanced data filtering pipeline for LLM classification. It has five steps: (1) label-based filtering keeps only training samples matching the target label; (2) confidence thresholding discards samples whose softmax confidence falls below a threshold such as 0.8; (3) context and feature extraction identifies words, phrases, or structural patterns that allow the LLM to label those samples correctly; (4) context aggregation and question formulation compresses these patterns into a small set of questions for GPT-4o; and (5) during testing, the LLM is prompted with those questions and must answer in a fixed two-line format with a confidence score. The machinery works by replacing open-ended image interpretation with a focused checklist, which simultaneously boosts accuracy and reduces the number of tokens the model must generate.

What would settle it

Take the chest X-ray dataset, hold out a fresh test split that is never shown to the filtering pipeline, rebuild the context questions using only the training split, and measure GPT-4o's accuracy; if it falls back to roughly 62% rather than 82%, the reported gain is caused by test information leaking into the prompt.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that CNN-based models remain the more accurate, better-calibrated, and far more efficient choice for medical image classification, while LLMs are overconfident and resource-hungry. The paper also demonstrates a concrete way to narrow the gap: an enhanced data filtering method that uses the training set to derive context questions, then injects those questions into the LLM prompt at test time. In the chest X-ray experiment this raises GPT-4o accuracy from 62% to 82.01%, keeps the average confidence score at 0.93, lowers average execution time from 6.23 seconds to 2.35 seconds, and reduces energy consumption from 1.84 to 1.65 watt-hours. The authors frame this as evidence that additional filtering on top of LLMs can yield substantial performance gains, and they anticipate similar gains across other LLMs and modalities.

Load-bearing premise

The central claim that filtering lifts LLM accuracy depends on the test set being truly unseen and not used to design the filter's context questions; the paper reports a training/validation/unknown split of 80/20/20, which sums to 120% and is internally inconsistent, so a leakage-free replication is the load-bearing premise.

Editorial extensions

If this is right

  • For all three modalities, CNNs are the safer default for real-time clinical screening, since they combine higher accuracy with lower latency, energy use, and CO2 emissions.
  • LLM-based diagnostic pipelines can be substantially improved by prompt filtering alone, without fine-tuning or architectural changes, at least on chest X-rays.
  • LLM confidence scores are poorly calibrated for medical image tasks, so any clinical use must either recalibrate them or abstain when confidence is unreliable.
  • The filtering approach reduces execution time and energy as well as improving accuracy, making it an efficiency intervention, not just a quality intervention.
  • The paper explicitly proposes hybrid models that merge CNNs with LLMs as future work, which would aim to combine visual accuracy with contextual reasoning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the filtering gain transfers to other LLMs and modalities, it would turn medical LLM classification into a prompt-engineering problem rather than a model-scaling problem; the paper only demonstrates the gain on chest X-rays with GPT-4o.
  • The reported split of 80% training, 20% validation, and 20% unknown sums to 120%, which suggests the partition is misdescribed; a replicator should first clarify how the test set was actually split and whether validation or test images informed the context questions.
  • The strong overconfidence of GPT-4o (0.93 confidence at 62% accuracy) suggests a simple rejection rule—refuse to answer when confidence is low—could improve safety, but the paper does not evaluate that.
  • Because the filtering questions are derived from training-set statistics, the method should be sensitive to distribution shift; checking whether the 82% accuracy holds on an external hospital dataset would be a natural stress test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The manuscript reports a comparative evaluation of two CNN-based classifiers and two multimodal LLMs (GPT-4o, Llama3.2-vision) on three public medical imaging datasets (chest X-ray, brain MRI, chest CT). Models are compared by accuracy, F1-score, average confidence, execution time, energy consumption, and estimated CO2 emissions. The main findings are that the CNNs outperform the LLMs on all three tasks, and that an 'enhanced data filtering' procedure raises GPT-4o's chest X-ray accuracy from 62% to 82% while reducing execution time and energy. The paper does not provide code, data splits, or prompts.

Significance. If the reported comparisons are reproducible, the paper would provide a useful practical benchmark, particularly the energy/carbon comparison and the attempt to improve LLM image classification through prompt-level filtering. The use of publicly available datasets and multiple evaluation metrics is a strength, and the CO2/energy analysis addresses an often-overlooked dimension. However, the absence of reproducibility artifacts and several internal inconsistencies currently prevent the results from supporting the abstract's claims at the advertised strength.

major comments (5)
  1. [Sec. 3.1, Dataset split] The split described as 'training (80%), validation (20%), and unknown (20%)' sums to 120% and is internally inconsistent. Because every reported test-set number depends on which images were held out, the authors must state the exact split (e.g., 80/10/10 or 60/20/20) and provide per-class counts for each subset.
  2. [Table 2 and Sec. 3.2] The headline filtering gain is not interpretable as stated. The w/o-filtering condition is not defined: the paper does not give the prompt, the label space, or the number of test samples used for the 62% figure. The filtered prompt in Fig. 2 asks a binary 'normal or COVID-affected' question, while the chest X-ray dataset contains four classes (Sec. 2.1). If the baseline used a four-way question or if Step 2's confidence thresholding removed test samples, the 82% result would reflect a changed task or test-set selection rather than the filtering method. Please report the exact prompts, label spaces, and per-condition test counts for both conditions.
  3. [Tables 1 and 2] No error bars, confidence intervals, or statistical tests are reported, and no information is given about the number of runs, random seeds, or API temperature settings. Given that the LLM accuracies for chest X-ray (0.65 vs 0.62) are close and API responses are stochastic, the observed differences may not be stable. The authors should add repeated-run statistics and significance tests.
  4. [Sec. 3.2, Step 2] The 0.8 confidence threshold is chosen from the training set, but no sensitivity analysis is provided. Because the main performance gain in Table 2 depends on this threshold, the authors should show how the result varies with the threshold and confirm that the threshold was selected without using the test set.
  5. [Sec. 4.3, Computational efficiency] The energy and CO2 estimates are not reproducible because the hardware, power measurement method, carbon intensity values, and formulas are not stated. For example, Table 2 reports energy in 'W-H' without specifying whether this is a measured per-query value or derived from runtime and a fixed power draw. Please provide the measurement details and uncertainty estimates.
minor comments (6)
  1. [Sec. 2.1 and Sec. 3.1] Sec. 2.1 lists four classes for the chest X-ray dataset, but Sec. 3.1 says the CNN is 'optimized to classify COVID-19 and normal chest X-rays' and Fig. 2 shows a binary prompt; please clarify which label space each experiment uses.
  2. [Sec. 4.2] The text states the CNN achieves 'precision, recall, accuracy, and F1-score of 0.83', but Table 1 reports only accuracy, F1-score, and confidence; precision and recall values are missing.
  3. [Sec. 3.1] The CNN baselines are referred to as 'COVID-Roentgen CNN' and 'TensorFlow-based CNN' without architecture details, hyperparameters, or training procedures; please add enough detail to reproduce the CNN baselines.
  4. [Table 2] The unit 'W-H' should be the standard symbol 'Wh' (watt-hour).
  5. [Abstract and Introduction] The framework is described as 'multimodal', but the LLM experiments appear to use image-only prompts plus aggregated textual context; please clarify what makes the framework multimodal.
  6. [Sec. 4.4] The statement 'We anticipate similar gains with other LLMs across various imaging modalities' is speculative; please either add those experiments or soften the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the benchmark results are empirical and the filter threshold is an experimental choice, not a fitted input that defines the outcome.

full rationale

The paper makes no formal derivation; it reports an empirical benchmark comparing CNNs and LLMs. The only candidates for circularity are (i) the confidence-threshold filtering in Sec. 3.2, where the threshold 0.8 is chosen from training-set softmax scores, and (ii) two self-citations ([10] and [24]). Neither is load-bearing. The threshold is an experimental design choice, and the filtered prompt is evaluated on test images, so the reported 62% to 82% accuracy gain is not forced by construction. The self-citations provide related-work context (LLM hallucination assessment and federated learning) and do not supply any premise used to derive the accuracy numbers. The paper's internal inconsistency that the split adds to 120% and the lack of a baseline prompt/label-space description in Table 2 are validity and comparability concerns, not circularity: no equation or fitted parameter is defined in terms of the target result. Therefore no circular step can be exhibited, and the score is 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities, forces, or conserved quantities. Its main load-bearing assumptions are about data integrity, confidence-score semantics, and the validity of energy estimates.

free parameters (1)
  • confidence threshold for filtering = 0.8
    The filtering method in Section 3.2 discards training samples with confidence below 0.8. The choice is not varied or justified, and no sensitivity analysis is reported, so it acts as a hand-picked parameter.
assumptions (3)
  • domain assumption The reported test subsets are completely unseen and were not used to construct the filtering questions.
    The paper's central accuracy claims depend on this. The inconsistent split statement (80/20/20) and absence of leakage checks make this assumption fragile.
  • domain assumption The LLM-provided confidence scores are comparable across models and reflect prediction reliability.
    Section 4.5 treats high average confidence as a sign of overconfidence, which only makes sense if the scores are meaningful uncertainty estimates to begin with.
  • domain assumption Energy consumption and CO2 estimates derived from execution time and published carbon intensity factors are accurate enough for the comparisons.
    Section 4.3 cites [26] but does not account for hardware differences, thermal effects, or measurement granularity, so the reported energy and CO2 numbers are approximate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Large Language Models Challenge CNNs in Medical Image Analysis?." pith.science (2026). https://pith.science/paper/7B4ZRODO

@misc{pith2026250523503,
  author       = {Pith},
  title        = {Pith review of: Can Large Language Models Challenge CNNs in Medical Image Analysis?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7B4ZRODO}},
  note         = {Machine review of arXiv:2505.23503}
}
abstract

This study presents a multimodal AI framework designed for precisely classifying medical diagnostic images. Utilizing publicly available datasets, the proposed system compares the strengths of convolutional neural networks (CNNs) and different large language models (LLMs). This in-depth comparative analysis highlights key differences in diagnostic performance, execution efficiency, and environmental impacts. Model evaluation was based on accuracy, F1-score, average execution time, average energy consumption, and estimated $CO_2$ emission. The findings indicate that although CNN-based models can outperform various multimodal techniques that incorporate both images and contextual information, applying additional filtering on top of LLMs can lead to substantial performance gains. These findings highlight the transformative potential of multimodal AI systems to enhance the reliability, efficiency, and scalability of medical diagnostics in clinical settings.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 40 canonical work pages

  1. [1]

    INTRODUCTION The rapid evolution of artificial intelligence (AI) has ushered in transformative approaches for medical diagnostics, particu- larly in the analysis of diagnostic imaging data [1]. Over the past decade, machine learning techniques have been applied successfully to tasks such as image classification, segmenta- tion, and anomaly detection in mo...

    work page Pith review arXiv 2025

  2. [2]

    DATASETS AND RELATED WORKS In this section, we detail the datasets utilized in this study and review existing literature relevant to the classification of medical images. 2.1. Dataset Selection To ensure a comprehensive evaluation, we select the follow- ing publicly available datasets covering a range of imaging modalities and diagnostic tasks. These data...

    work page

  3. [3]

    ANALYSIS OF THE MODELS In this section, we present the various models employed in this study for classifying medical images, followed by an analysis of the corresponding classification processes. 3.1. Classification Approaches In this study, we investigate two complementary paradigms for medical image classification: large language models (LLMs) and deep ...

    work page

  4. [4]

    Context Aggregation and Question Formulation: We aggregate these contextual insights from all filtered sam- ples and present them to GPT-4o. This step produces a consolidated understanding of the features driving clas- sification decisions, as well as a set of targeted questions whose answers are critical for label determination

    work page

  5. [5]

    From this dataset, we first extract only those samples correspond- ing to the desired label (e.g., “normal” in the context of COVID classification)

    Initial Label-Based Filtering: We begin by enumerat- ing the responses across the entire training set. From this dataset, we first extract only those samples correspond- ing to the desired label (e.g., “normal” in the context of COVID classification)

    work page

  6. [6]

    Confidence Thresholding: To further refine the dataset, we associate each sample with a confidence score (such as the output of a softmax layer reflecting the model’s certainty in its prediction). We then discard samples with scores below a chosen threshold (e.g., 0.8), thereby retaining only high-confidence, label-consistent samples that are more likely ...

    work page

  7. [7]

    Context and Feature Extraction: Next, we examine these high-quality samples to identify key contextual elements and features (e.g., specific words, phrases, or structural patterns) that enable the Large Language Model (LLM) to classify them as the chosen label

    work page

  8. [8]

    Large language models in medicine,

    Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting, “Large language models in medicine,” Nature medicine, vol. 29, no. 8, pp. 1930– 1940, 2023

    work page 1930

Show all 41 references
  1. [9]

    Application During Testing: Finally, in the testing phase, we incorporate these targeted questions into the user prompt, rather than attempting to enumerate all possible features. By prompting the LLM to respond to the key questions directly, we streamline the classifica- tion...

  2. [10]

    Normal Heart and Mediastinum

  3. [11]

    Intact Bony Structures

  4. [12]

    The first line should contain only either normal or abnormal

    Empty Pleural Spaces Based on the above criteria, does this image appear normal or COVID-affected? Please reply in exactly two lines. The first line should contain only either normal or abnormal. The second line should be in the format Confidence Score: <number> where <number>...

  5. [13]

    Evaluation Metrics The performance of CNNs and LLMs was evaluated using sev- eral key metrics

    RESULTS AND DISCUSSIONS 4.1. Evaluation Metrics The performance of CNNs and LLMs was evaluated using sev- eral key metrics. Accuracy is measured by the percentage of correctly classified samples, providing an overall performance indicator. Precision assessed the proportion of ...

  6. [14]

    CONCLUSION AND FUTURE WORK This study compared CNNs and LLMs for medical image clas- sification, highlighting their strengths and limitations. CNNs demonstrated superior accuracy and efficiency, making them well-suited for real-time clinical applications, while LLMs showed pot...

  7. [15]

    A review of deep learning in medical imag- ing: Imaging traits, technology trends, case studies with progress highlights, and future promises,

    S. Kevin Zhou, Hayit Greenspan, Christos Davatzikos, James S. Duncan, Bram Van Ginneken, Anant Madab- hushi, et al., “A review of deep learning in medical imag- ing: Imaging traits, technology trends, case studies with progress highlights, and future promises,” Proceedings of ...

    2021

  8. [16]

    This dataset is commonly used to benchmark AI models for classifying chest X-rays

    contains X-ray images labeled as COVID-19, pneumonia, lung opacity, and normal. This dataset is commonly used to benchmark AI models for classifying chest X-rays. [17]. Brain Tumor MRI Dataset: The Brain Tumor MRI Dataset [18], available on Kaggle, includes T1-weighted MRI ima...

  9. [17]

    Two-stage selective ensemble of CNN via deep tree train- ing for medical image classification,

    Yun Yang, Yuanyuan Hu, Xingyi Zhang, and Song Wang, “Two-stage selective ensemble of CNN via deep tree train- ing for medical image classification,” IEEE Transactions on Cybernetics, vol. 52, no. 9, pp. 9194–9207, 2022

    2022

  10. [18]

    Blockchain- federated-learning and deep learning models for COVID- 19 detection using CT imaging,

    Rajesh Kumar, Abdullah Aman Khan, Jay Kumar, Za- kria, Noorbakhsh Amiri Golilarz, Simin Zhang, Yang Ting, Chengyu Zheng, and Wenyong Wang, “Blockchain- federated-learning and deep learning models for COVID- 19 detection using CT imaging,” IEEE Sensors Journal, vol. 21, no. 14,...

    2021

  11. [19]

    Znet: Deep learning approach for 2D MRI brain tumor segmentation,

    Mohammad Ashraf Ottom, Hanif Abdul Rahman, and Ivo D. Dinov, “Znet: Deep learning approach for 2D MRI brain tumor segmentation,” IEEE Jour. Transl. Eng. Health Med, vol. 10, pp. 1–8, 2022

    2022

  12. [20]

    UNETR: Transformers for 3D medical image segmentation,

    Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu, “UNETR: Transformers for 3D medical image segmentation,” in Proceedings of the IEEE/CVF winter conference on applications of com- puter vision, 2022, pp. 574–584

    2022

  13. [21]

    CapsCovNet: A modified capsule network to diagnose COVID-19 from multimodal medical imaging,

    A. F. M. Saif, Tamjid Imtiaz, Shahriar Rifat, Celia Shah- naz, Wei-Ping Zhu, and M. Omair Ahmad, “CapsCovNet: A modified capsule network to diagnose COVID-19 from multimodal medical imaging,” IEEE Transactions on Artificial Intelligence, vol. 2, no. 6, pp. 608–617, 2021

    2021

  14. [22]

    Unsupervised medical image translation with adversarial diffusion models,

    Muzaffer Özbey, Onat Dalmaz, Salman U. H. Dar, Hasan A. Bedel, ¸ Saban Özturk, Alper Güngör, and Tolga Çukur, “Unsupervised medical image translation with adversarial diffusion models,” IEEE Transactions on Medical Imaging, vol. 42, no. 12, pp. 3524–3539, 2023

    2023

  15. [23]

    Eval- uation and mitigation of the limitations of large language models in clinical decision-making,

    Paul Hager, Friederike Jungmann, Robbie Holland, Ku- nal Bhagat, Inga Hubrecht, Manuel Knauer, et al., “Eval- uation and mitigation of the limitations of large language models in clinical decision-making,” Nature medicine, vol. 30, pp. 2613–2622, 2024

    2024

  16. [24]

    Hallucinations and key information extraction in medical texts: A comprehensive assess- ment of open-source large language models,

    Anindya Bijoy Das, Shibbir Ahmed, and Shah- newaz Karim Sakib, “Hallucinations and key information extraction in medical texts: A comprehensive assess- ment of open-source large language models,” preprint, arXiv:2504.19061, 2025

    2025 arXiv

  17. [25]

    Performance of ChatGPT on USMLE: potential for AI-assisted medical education us- ing large language models,

    Tiffany H Kung, Morgan Cheatham, Arielle Medenilla, Czarina Sillos, et al., “Performance of ChatGPT on USMLE: potential for AI-assisted medical education us- ing large language models,” PLoS digital health, 2023

    2023

  18. [26]

    Explainability for large language models: A survey,

    Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du, “Explainability for large language models: A survey,” ACM Transactions on Intelligent Systems and Technology, vol. 15, no. 2, pp. 1–38, 2024

    2024

  19. [27]

    CXR-LLaV A: a multimodal large language model for interpreting chest X-ray images,

    Seowoo Lee, Jiwon Youn, Hyungjin Kim, Mansu Kim, and Soon Ho Yoon, “CXR-LLaV A: a multimodal large language model for interpreting chest X-ray images,” Eu- ropean Radiology, pp. 1–13, 2025

    2025

  20. [28]

    A LLM-based hybrid-transformer diagnosis system in healthcare,

    Dongyuan Wu, Liming Nie, Rao Asad Mumtaz, and Kadambri Agarwal, “A LLM-based hybrid-transformer diagnosis system in healthcare,”IEEE Journal of Biomed- ical and Health Informatics, pp. 1–12, 2024

    2024

  21. [29]

    Chat- CAD+: Toward a universal and reliable interactive CAD using LLMs,

    Zihao Zhao, Sheng Wang, Jinchen Gu, Yitao Zhu, Lanzhuju Mei, Zixu Zhuang, Zhiming Cui, et al., “Chat- CAD+: Toward a universal and reliable interactive CAD using LLMs,” IEEE Transactions on Medical Imaging, vol. 43, no. 11, pp. 3755–3766, 2024

    2024

  22. [30]

    Can AI help in screening viral and COVID-19 pneumo- nia?,

    Muhammad EH Chowdhury, Tawsifur Rahman, Amith Khandakar, Rashid Mazhar, Muhammad Abdul Kadir, Zaid Bin Mahbub, Khandakar Reajul Islam, Muham- mad Salman Khan, Atif Iqbal, Nasser Al Emadi, et al., “Can AI help in screening viral and COVID-19 pneumo- nia?,” IEEE Access, vol. 8, ...

    2020

  23. [31]

    Exploring the effect of image enhancement techniques on COVID-19 detec- tion using chest X-ray images,

    Tawsifur Rahman, Amith Khandakar, Yazan Qiblawey, Anas Tahir, Serkan Kiranyaz, et al., “Exploring the effect of image enhancement techniques on COVID-19 detec- tion using chest X-ray images,” Computers in biology and medicine, vol. 132, pp. 104319, 2021

    2021

  24. [32]

    Brain tumor MRI dataset,

    Masoud Nickparvar, “Brain tumor MRI dataset,” Kaggle, 2020, https://www.kaggle.com/datasets/masoud nickparvar/brain-tumor-mri-dataset

    2020

  25. [33]

    Chest CT-scan images dataset,

    Mohamed Hany, “Chest CT-scan images dataset,” Kag- gle, 2021, https://www.kaggle.com/datasets/moha medhanyyy/chest-ctscan-images

    2021

  26. [34]

    MMBERT: Mul- timodal BERT pretraining for improved medical VQA,

    Yash Khare, Viraj Bagal, Minesh Mathew, Adithi Devi, U Deva Priyakumar, and CV Jawahar, “MMBERT: Mul- timodal BERT pretraining for improved medical VQA,” in 18th International Symposium on Biomedical Imaging (ISBI). IEEE, 2021, pp. 1033–1036

    2021

  27. [35]

    Multi-task paired masking with alignment modeling for medical vision-language pre-training,

    Ke Zhang, Yan Yang, Jun Yu, Hanliang Jiang, Jianping Fan, Qingming Huang, and Weidong Han, “Multi-task paired masking with alignment modeling for medical vision-language pre-training,” IEEE Transactions on Multimedia, vol. 26, pp. 4706–4721, 2024

    2024

  28. [36]

    On large visual language models for medical imaging analy- sis: An empirical study,

    Minh–Hao Van, Prateek Verma, and Xintao Wu, “On large visual language models for medical imaging analy- sis: An empirical study,” in IEEE/ACM Conference on Connected Health: Applications, Systems and Engineer- ing Technologies (CHASE), 2024, pp. 172–176

    2024

  29. [37]

    Large language mod- els in healthcare and medical domain: A review,

    Zabir Al Nazi and Wei Peng, “Large language mod- els in healthcare and medical domain: A review,” in Informatics. MDPI, 2024, vol. 11, p. 57

    2024

  30. [38]

    Ex- plainable vertical federated learning for healthcare: En- suring privacy and optimal accuracy,

    Shahnewaz Karim Sakib and Anindya Bijoy Das, “Ex- plainable vertical federated learning for healthcare: En- suring privacy and optimal accuracy,” in IEEE Inter- national Conference on Big Data (BigData), 2024, pp. 5068–5077

    2024

  31. [39]

    On calibration of modern neural networks,

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Wein- berger, “On calibration of modern neural networks,” in International Conference on Machine Learning (ICML), 2017, pp. 1321–1330

    2017

  32. [40]

    Estimating the carbon footprint of bloom, a 176b parameter language model,

    Alexandra Sasha Luccioni, Sylvain Viguier, and Anne- Laure Ligozat, “Estimating the carbon footprint of bloom, a 176b parameter language model,” J. Mach. Learn. Res., vol. 24, no. 1, Jan. 2023

    2023

  33. [41]

    Expert-level detection of pathologies from unannotated chest x-ray images via self-supervised learning,

    Ekin Tiu, Ellie Talius, Pujan Patel, et al., “Expert-level detection of pathologies from unannotated chest x-ray images via self-supervised learning,” Nature Biomedical Engineering, vol. 6, pp. 1399–1406, 2022

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.