Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

MCA-RG: Enhancing LLMs with Medical Concept Alignment for Radiology Report Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Aligning chest X-ray visual features to separate anatomy and pathology concepts produces reports that top prior systems on two public benchmarks.

desk verdict A credible incremental RRG method whose headline SOTA claim is undercut by a mixed evaluation protocol and missing artifacts. read the letter →

arxiv 2507.06992 v2 pith:H7DMM44G submitted 2025-07-09 cs.CV cs.AI

Qilong Xing , Zikai Song , Youjia Zhang , Na Feng , Junqing Yu , Wei Yang This is my paper
classification cs.CVcs.AI
keywords radiologyreportgenerationmedicalconceptalignmentchestX-raylargelanguagemodelscontrastivelearningfeaturegatingclinicalefficacybanks
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 tries to establish that radiology report generation improves when the visual features passed to a large language model are separated by medical concept instead of pooled into one global image representation. The authors build a pathology bank and an anatomy bank from report text, align chest X-ray image regions with each concept, then enhance and gate those concept-specific features before report generation. On the MIMIC-CXR and CheXpert Plus benchmarks they report top scores in both language metrics and clinical metrics, including an example-based clinical F1 of 0.408 and a macro F1 of 0.335 on MIMIC-CXR, where clinical F1 measures overlap with the ground-truth finding labels. If these results hold, the practical message is that explicit concept alignment is a stronger route to clinically accurate reports than feeding a model a single conflated feature vector.

What carries the argument

The central mechanism is concept-aligned visual feature extraction: report entities extracted by a named-entity relation tool are grouped into pathology and anatomy banks, and each banked concept acts as a query in a Transformer decoder over image features. A contrastive loss pulls anatomical features for the same structure closer across patients, and a matching loss raises similarity between a present pathology and the anatomy where it is found. A feature gate computes the entropy of each concept’s multi-head attention map and applies a learned projection followed by a sigmoid to suppress low-quality concept features. The large language model is conditioned on the gated concept features through separate projection modules for pathology and anatomy.

What would settle it

Take a trained model and, at inference, replace each concept’s multi-head attention map with a uniform distribution while holding all other inputs fixed: the stated rationale predicts the gating value should fall and report quality should drop; if neither happens, the entropy assumption is not the operative mechanism.

Watch

Extended reading notes

Core claim

The central claim is that conflated visual features are the bottleneck: pathologies and anatomies mapped through the same representation interfere with each other, causing wrong findings and wrong locations. MCA-RG’s fix is to use text-derived concept banks as queries in attention modules, producing one visual feature per medical concept, enrich those features with concept descriptions, refine anatomy features by contrastive learning across patients and pathology features by a matching loss that ties each present pathology to its anatomical location, and then gate each concept’s features by a learned projection of attention entropy. The final large language model, fine-tuned with low-rank adapters, receives these concept-specific features rather than one global vector. The paper reports that this design attains a BLEU-4 of 0.128, a METEOR of 0.163, a ROUGE-L of 0.300, and example-based and macro clinical F1 of 0.408 and 0.335 on MIMIC-CXR, the top values among the compared methods, with analogous gains on CheXpert Plus, and a case study in which the model identifies pneumonia in the right lower lung while avoiding the false positives a baseline generates.

Load-bearing premise

The gating mechanism’s stated rationale assumes that a concept’s attention map becoming more uniform means its visual features are less informative; if that premise is wrong, the entropy-based explanation is wrong even if the learned gate still helps.

Editorial extensions

If this is right

  • The ablation ties the largest single jump, 10.6% average improvement in natural-language metrics and a macro F1 rise from 0.205 to 0.288, to the two alignment losses, indicating feature alignment itself drives most of the gain.
  • Adding anatomy contrastive learning and the pathology-anatomy matching loss together raises the average NLG gain to 18.6% and macro F1 to 0.324.
  • Adding the feature gate brings the final average NLG gain to 20.5% and macro F1 to 0.335, so gating contributes beyond alignment and enhancement.
  • The concept banks built once from MIMIC-CXR reports are reused on CheXpert Plus with top reported scores, showing the concept vocabulary transfers across datasets.
  • The attention visualizations on the pneumonia and pneumothorax localization datasets indicate that attention for a term such as pneumonia lands on disease-relevant regions rather than unrelated areas.

Reading between the lines

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

  • The same recipe should transfer to other structured reporting domains, such as pathology, ultrasound, or retinal imaging, wherever reports name a finding and a location; a test would be building banks from those reports and measuring clinical F1.
  • The entropy-gating rationale is separable from the learned projection, so an ablation that removes the entropy input while keeping the gate would reveal whether the stated uniform-attention premise is actually the mechanism.
  • Automated clinical F1 is a label-overlap measure, so a radiologist reading study would be the next natural test of whether the concept-aligned reports are more clinically usable.
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

4 major / 5 minor

Summary. The paper proposes MCA-RG, a radiology report generation framework that constructs pathology and anatomy concept banks from RadGraph entities, enriches them with GPT-4, aligns image features to these concepts with transformer decoders, applies anatomy-based contrastive learning, a pathology--anatomy matching loss, and a learned feature gating mechanism, and then feeds the gated concept features to a LoRA-tuned LLaMA2 to generate reports. It reports experiments on MIMIC-CXR and CheXpert Plus, including ablations on MIMIC-CXR, and claims superior performance over prior RRG methods.

Significance. If the empirical claims hold, the paper offers a reasonably complete recipe for concept-aware RRG: the loss terms are well-defined, the conceptual separation of pathology and anatomy features is intuitive, and Table 3 shows each proposed component contributes to the reported improvement. The CheXpert Plus comparison, where all baselines are marked as re-implementations under the same protocol, is more credible than the MIMIC-CXR tables. The main weaknesses are the mixed evaluation protocols behind the MIMIC-CXR SOTA claim, the absence of variance or significance information, and the under-specified label construction; these are all fixable within a revision.

major comments (4)
  1. [Tables 1–2] The headline claim that MCA-RG achieves the highest reported BLEU-4, METEOR, ROUGE-L, and clinical F1 on MIMIC-CXR is not supported by a single evaluation protocol. Table 1 states that most MIMIC-CXR baseline scores are cited from original papers and [18,28], while only MiniGPT-Med* and Ours are computed in-house; Table 2 similarly cites clinical F1 scores from [20,28] and official GitHub sources. Since BLEU/METEOR/ROUGE depend on report preprocessing, tokenization, and split selection, and clinical F1 depends on the labeler and label set, the reported margins (e.g., BLEU-4 0.128 vs WarmStart 0.124; macro F1 0.335 vs 0.306) cannot be attributed to the method. The authors should re-run all baselines on MIMIC-CXR under an identical preprocessing and evaluation pipeline, as they do for CheXpert Plus, or explicitly restrict the SOTA claim to the re-implemented comparisons.
  2. [§3.2, Tables 1–3] No variance or significance information is reported for any experimental result. The margins over the strongest baselines are small relative to typical run-to-run variation in LLM fine-tuning, so even under a unified protocol the current tables do not establish that the differences are real. Report mean and standard deviation over at least three seeds, or bootstrap confidence intervals, for the main and ablation results.
  3. [§2.2, Eq. (3)] The anatomy and pathology labels Y^a and Y^p are central to the alignment and matching losses, but the manuscript never states how they are constructed. It says only that RadGraph extracts triplets; the mapping from triplets to binary healthy/pathology labels and to the E_ij matrix in Eq. (5) needs to be defined precisely. Without this, the method is not reproducible and the effect of the label construction on the reported gains cannot be assessed.
  4. [§2.3, Eq. (6)] The feature gating mechanism is motivated by the claim that attention maps with a more uniform value distribution indicate low-quality concept features. This premise is not validated independently; since the gating projection W^{g,i} is learned, the gate could improve performance even if the entropy hypothesis were false, and then the explanation would be incorrect. Please provide direct evidence (e.g., correlation of gate outputs with detection errors or a controlled experiment with a non-entropy gate) to support the stated rationale, or revise the motivation accordingly.
minor comments (5)
  1. [§2.2, Eq. (4)] The notation for v'^{a}_{j,k0} and the denominator index is ambiguous; clarify which sample k0 selects and that the positive pair is the same anatomy across samples.
  2. [§2.1] The use of GPT-4 for concept enrichment is not reproducible without the prompt, version, and sampling settings; please include these details or replace with an open model.
  3. [§3.3, Fig. 4] The attention visualization on RSNA and SIIM-ACR is only qualitative; no quantitative localization metric is provided.
  4. [§3.1] The choices β0=0.5 and β1=0.3 and the concept bank sizes 51/67 are not ablated; a brief sensitivity analysis would help.
  5. [Table 3] The AVG.Δ column reports relative improvements, but the caption does not state whether these are arithmetic means of per-metric relative gains; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MCA-RG's losses are standard supervised objectives trained on external annotations, not fitted quantities renamed as predictions.

full rationale

The paper's derivation chain is conventional supervised learning. Concept banks are built by applying RadGraph to training reports, and the alignment, contrastive, matching, and generation losses (Eqs. 3-8) are supervised by labels Y_a, Y_p, E_ij, and the reference report text. These are training annotations, not quantities that reappear as the reported evaluation targets by construction: clinical F1 is computed against held-out test report entities, and NLG metrics compare generated text to test references. The feature-gating hypothesis about attention entropy is an unvalidated motivation, but the gate projection is learned end-to-end and ablated, so it is not a fitted parameter renamed as a prediction. The comparison of cited baseline scores from different papers is an evaluation-protocol fairness concern, not circularity. I found no load-bearing self-citation, no imported uniqueness theorem, and no renaming of a known empirical pattern as a derivation. Accordingly, the paper does not exhibit definitional or self-citation circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method's central claim rests on several unvalidated domain assumptions about concept extraction, knowledge enrichment, and feature quality. The free parameters are chosen by hand and no sensitivity analysis is provided.

free parameters (3)
  • Loss weights beta0 and beta1 = 0.5, 0.3
    Hand-chosen in Section 3.1 to balance auxiliary losses; no sensitivity analysis provided, yet they affect all reported results.
  • Concept bank sizes = 51 anatomical, 67 pathological
    Selected by frequency in MIMIC-CXR reports; the threshold is arbitrary and the same banks are reused for CheXpert Plus without verification.
  • LoRA rank and alpha = 64, 64
    Set in Section 3.1 for fine-tuning LLaMA2; standard defaults but still model capacity choices.
assumptions (5)
  • domain assumption RadGraph NER reliably extracts pathology and anatomy concepts and E_ij relations from reports (Eq. 1).
    All concept banks, matching labels, and supervision signals derive from RadGraph output; errors propagate into training signals.
  • domain assumption GPT-4 enrichment adds useful medical knowledge for concept embeddings.
    Section 2.1 states concepts are enriched with GPT-4 details; no ablation or evaluation of this step is provided.
  • ad hoc to paper Attention-map entropy correlates with visual feature quality.
    Section 2.3 asserts uniform attention indicates low-quality features; this premise is not empirically validated.
  • domain assumption Contrastive learning on anatomical features across samples improves generalization.
    Standard contrastive assumption, but the paper does not analyze failure cases or measure feature separation quantitatively beyond one illustrative figure.
  • domain assumption Concept banks from MIMIC-CXR cover CheXpert Plus vocabulary.
    Stated in Section 3.1 without checking CheXpert Plus report frequencies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MCA-RG: Enhancing LLMs with Medical Concept Alignment for Radiology Report Generation." pith.science (2026). https://pith.science/paper/H7DMM44G

@misc{pith2026250706992,
  author       = {Pith},
  title        = {Pith review of: MCA-RG: Enhancing LLMs with Medical Concept Alignment for Radiology Report Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H7DMM44G}},
  note         = {Machine review of arXiv:2507.06992}
}
read the original abstract

Despite significant advancements in adapting Large Language Models (LLMs) for radiology report generation (RRG), clinical adoption remains challenging due to difficulties in accurately mapping pathological and anatomical features to their corresponding text descriptions. Additionally, semantic agnostic feature extraction further hampers the generation of accurate diagnostic reports. To address these challenges, we introduce Medical Concept Aligned Radiology Report Generation (MCA-RG), a knowledge-driven framework that explicitly aligns visual features with distinct medical concepts to enhance the report generation process. MCA-RG utilizes two curated concept banks: a pathology bank containing lesion-related knowledge, and an anatomy bank with anatomical descriptions. The visual features are aligned with these medical concepts and undergo tailored enhancement. We further propose an anatomy-based contrastive learning procedure to improve the generalization of anatomical features, coupled with a matching loss for pathological features to prioritize clinically relevant regions. Additionally, a feature gating mechanism is employed to filter out low-quality concept features. Finally, the visual features are corresponding to individual medical concepts, and are leveraged to guide the report generation process. Experiments on two public benchmarks (MIMIC-CXR and CheXpert Plus) demonstrate that MCA-RG achieves superior performance, highlighting its effectiveness in radiology report generation.

Figures

Figures reproduced from arXiv: 2507.06992 by the authors.

Figure 1
Figure 1. Our MCA-RG first extracts and enriches the pathological and anatomical con￾cepts with medical knowledge, and then align the image features with concepts by guiding the model to diagnose pathologies and anatomical abnormalities. Anatomical features are enhanced through anatomy-based contrastive learning, while pathological features are refined using matching loss. Finally, a gating mechanism filters the visual featur… view at source ↗
Figure 2
Figure 2. An illustration comparing the reports generated by different models for a given input radiograph. Each matched descriptions are highlighted using consistent colors to facilitate comparison. Red background indicates wrong descriptions. highest ROUGE-L scores of 0.3 and 0.266 on two benchmarks. In clinical accu￾racy ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. The visualization of inter class fea￾ture distance. Ground Truth WarmStart MiniGPT-Med Ours Pneumonia Pneumothorax 1.0 0.5 0.0 Ground Truth WarmStart MiniGPT￾Med Ours Pneumonia Pneumothorax 1.0 0.5 0.0 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. R2GenKG: Hierarchical Multi-modal Knowledge Graph for LLM-based Radiology Report Generation

    cs.CV 2025-08 reject novelty 5.0 of 10

    R2GenKG generates X-ray reports with an LLM conditioned on a GPT-4o-built multi-modal knowledge graph, reporting small metric gains on IU-Xray and CheXpert Plus.

Reference graph

Works this paper leans on

31 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [1]

    arXiv preprint arXiv:2307.01473 (2023)

    Ahmadi, R., Rajabi, M.J., Khalooie, M., Sabokrou, M.: Mitigating bias: En- hancing image classification by improving model explanations. arXiv preprint arXiv:2307.01473 (2023)

    arXiv 2023

  2. [2]

    arXiv preprint arXiv:2407.04106 (2024)

    Alkhaldi,A.,Alnajim,R.,Alabdullatef,L.,Alyahya,R.,Chen,J.,Zhu,D.,Alsinan, A., Elhoseiny, M.: Minigpt-med: Large language model as a general interface for radiology diagnosis. arXiv preprint arXiv:2407.04106 (2024)

    arXiv 2024

  3. [3]

    arXiv preprint arXiv:1904.03323 (2019)

    Alsentzer, E., Murphy, J.R., Boag, W., Weng, W.H., Jin, D., Naumann, T., McDermott, M.: Publicly available clinical bert embeddings. arXiv preprint arXiv:1904.03323 (2019)

    arXiv 2019

  4. [4]

    In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summa- rization

    Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with im- proved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summa- rization. pp. 65–72 (2005)

    2005

  5. [5]

    In: European conference on computer vision

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End- to-end object detection with transformers. In: European conference on computer vision. pp. 213–229. Springer (2020)

    2020

  6. [6]

    Chambon, P., Delbrouck, J.B., Sounack, T., Huang, S.C., Chen, Z., Varma, M., Truong, S.Q., Chuong, C.T., Langlotz, C.P.: Chexpert plus: Augmenting a large chest x-ray dataset with text radiology reports, patient demographics and addi- tional image formats (2024), https://arxiv.org/abs/2405.19538

    arXiv 2024

  7. [7]

    arXiv preprint arXiv:2303.09117 (2023)

    Chen, W., Liu, Y., Wang, C., Zhu, J., Zhao, S., Li, G., Liu, C.L., Lin, L.: Cross-modal causal intervention for medical report generation. arXiv preprint arXiv:2303.09117 (2023)

    arXiv 2023

  8. [8]

    In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Chen, W., Shen, L., Lin, J., Luo, J., Li, X., Yuan, Y.: Fine-grained image-text alignment in medical imaging enables explainable cyclic image-report generation. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 9494–9509 (2024)

    work page 2024

Show all 31 references
  1. [9]

    arXiv preprint arXiv:2204.13258 (2022)

    Chen, Z., Shen, Y., Song, Y., Wan, X.: Cross-modal memory networks for radiology report generation. arXiv preprint arXiv:2204.13258 (2022)

    2022 arXiv

  2. [10]

    arXiv preprint arXiv:2010.16056 (2020)

    Chen, Z., Song, Y., Chang, T.H., Wan, X.: Generating radiology reports via memory-driven transformer. arXiv preprint arXiv:2010.16056 (2020)

    2020 arXiv

  3. [11]

    arXiv preprint arXiv:2411.13025 (2024)

    Gu, T., Yang, K., An, X., Feng, Z., Liu, D., Cai, W.: Orid: Organ-regional information driven framework for radiology report generation. arXiv preprint arXiv:2411.13025 (2024)

    2024 arXiv

  4. [12]

    arXiv preprint arXiv:2404.15127 (2024) 10 Q

    He, S., Nie, Y., Chen, Z., Cai, Z., Wang, H., Yang, S., Chen, H.: Meddr: Diagnosis-guided bootstrapping for large-scale medical vision-language learning. arXiv preprint arXiv:2404.15127 (2024) 10 Q. Xing et al

    2024 arXiv

  5. [13]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Huang, Z., Zhang, X., Zhang, S.: Kiut: Knowledge-injected u-transformer for ra- diology report generation. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 19809–19818 (2023)

    2023

  6. [14]

    arXiv preprint arXiv:2106.14463 (2021)

    Jain, S., Agrawal, A., Saporta, A., Truong, S.Q., Duong, D.N., Bui, T., Chambon, P., Zhang, Y., Lungren, M.P., Ng, A.Y., et al.: Radgraph: Extracting clinical enti- ties and relations from radiology reports. arXiv preprint arXiv:2106.14463 (2021)

    2021 arXiv

  7. [15]

    arXiv preprint arXiv:2004.12274 (2020)

    Jing, B., Wang, Z., Xing, E.: Show, describe and conclude: On exploiting the struc- ture information of chest x-ray reports. arXiv preprint arXiv:2004.12274 (2020)

    2020 arXiv

  8. [16]

    Scientific data 6(1), 317 (2019)

    Johnson, A.E., Pollard, T.J., Berkowitz, S.J., Greenbaum, N.R., Lungren, M.P., Deng, C.y., Mark, R.G., Horng, S.: Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data 6(1), 317 (2019)

    2019

  9. [17]

    In: Text sum- marization branches out

    Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Text sum- marization branches out. pp. 74–81 (2004)

    2004

  10. [18]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Liu, C., Tian, Y., Chen, W., Song, Y., Zhang, Y.: Bootstrapping large language models for radiology report generation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 18635–18643 (2024)

    2024

  11. [19]

    arXiv preprint arXiv:2106.06965 (2021)

    Liu, F., Yin, C., Wu, X., Ge, S., Zou, Y., Zhang, P., Sun, X.: Contrastive atten- tion for automatic chest x-ray report generation. arXiv preprint arXiv:2106.06965 (2021)

    2021

  12. [20]

    Artificial intelligence in medicine144, 102633 (2023)

    Nicolson, A., Dowling, J., Koopman, B.: Improving chest x-ray report generation by leveraging warm starting. Artificial intelligence in medicine144, 102633 (2023)

    2023

  13. [21]

    arXiv preprint arXiv:2102.09777 (2021)

    Nooralahzadeh, F., Gonzalez, N.P., Frauenfelder, T., Fujimoto, K., Krauthammer, M.: Progressive transformer-based generation of radiology reports. arXiv preprint arXiv:2102.09777 (2021)

    2021 arXiv

  14. [22]

    In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics

    Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311–318 (2002)

    2002

  15. [23]

    In: 2015 IEEE 12th international symposium on biomedical imaging (ISBI)

    Roth, H.R., Lee, C.T., Shin, H.C., Seff, A., Kim, L., Yao, J., Lu, L., Summers, R.M.: Anatomy-specific classification of medical images using deep convolutional nets. In: 2015 IEEE 12th international symposium on biomedical imaging (ISBI). pp. 101–104. IEEE (2015)

    2015

  16. [24]

    Radiology: Artificial Intelligence1(1), e180041 (2019)

    Shih, G., Wu, C.C., Halabi, S.S., Kohli, M.D., Prevedello, L.M., Cook, T.S., Sharma, A., Amorosa, J.K., Arteaga, V., Galperin-Aizenberg, M., et al.: Aug- menting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia. Radiology...

    2019

  17. [25]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Tanida,T.,Müller,P.,Kaissis,G.,Rueckert,D.:Interactiveandexplainableregion- guided radiology report generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7433–7442 (2023)

    2023

  18. [26]

    arXiv preprint arXiv:2306.07971 (2023)

    Thawkar, O., Shaker, A., Mullappilly, S.S., Cholakkal, H., Anwer, R.M., Khan, S., Laaksonen, J., Khan, F.S.: Xraygpt: Chest radiographs summarization using medical vision-language models. arXiv preprint arXiv:2306.07971 (2023)

    2023 arXiv

  19. [27]

    arXiv preprint arXiv:2307.09288 (2023)

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bash- lykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  20. [28]

    arXiv preprint arXiv:2407.15158 (2024)

    Wang, F., Du, S., Yu, L.: Hergen: Elevating radiology report generation with lon- gitudinal data. arXiv preprint arXiv:2407.15158 (2024)

    2024 arXiv

  21. [29]

    arXiv preprint arXiv:2210.10163 (2022) Title Suppressed Due to Excessive Length 11

    Wang, Z., Wu, Z., Agarwal, D., Sun, J.: Medclip: Contrastive learning from un- paired medical images and text. arXiv preprint arXiv:2210.10163 (2022) Title Suppressed Due to Excessive Length 11

    2022 arXiv

  22. [30]

    arXiv preprint arXiv:2108.00316 (2021)

    Wu, J.T., Agu, N.N., Lourentzou, I., Sharma, A., Paguio, J.A., Yao, J.S., Dee, E.C., Mitchell, W., Kashyap, S., Giovannini, A., et al.: Chest imagenome dataset for clinical reasoning. arXiv preprint arXiv:2108.00316 (2021)

    2021 arXiv

  23. [31]

    Zawacki, A., Wu, C., Shih, G., Elliott, J., Fomitchev, M., Hussain, M., Lakhani, P., Culliton, P., Bao, S.: Siim-acr pneumothorax segmentation (2019)

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.