Pith. sign in

REVIEW 5 major objections 5 minor 27 references

A Generative Framework for Bidirectional Image-Report Understanding in Chest Radiography

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

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

Pith's one-line read MAViLT tops prior models on chest X-ray report, image, and VQA

desk verdict The architecture is plausible but the empirical core collapses: the generalization section cites tables that do not contain its own numbers, and the baselines are mis-cited, so the SOTA claim is unverifiable as written. read the letter →

arxiv 2502.05926 v1 pith:YYMYT7TN submitted 2025-02-09 eess.IV cs.CLcs.CV

Nicholas Evans , Stephen Baker , Miles Reed This is my paper · ORCID
classification eess.IVcs.CLcs.CV
keywords chestX-rayradiologyreportgenerationmedicalimagesynthesisvisualquestionansweringvision-languagemodelVQ-GANclinicalgradient-weightedlosshierarchicalfine-tuning
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

The paper sets out to show that a single large language model, extended to see and generate images, can handle the two directions of chest X-ray understanding at once: turning a CXR into a radiology report and turning a report back into a CXR, with clinical question answering as a third task. The proposed framework, MAViLT, adds a clinical gradient-weighted loss to VQ-GAN tokenization so that lesion boundaries and texture survive compression, and schedules training in two stages to align vision and language without forgetting the LLM's pretrained abilities. On MIMIC-CXR and Indiana CXR it reports higher BLEU, ROUGE-L, METEOR, and CIDEr for report generation, a lower FID for image synthesis, and higher accuracy and AUROC for VQA than four prior systems, with radiologist ratings favoring its output quality. If these results hold, a single model could assist clinical workflows that currently need separate report generators, image synthesizers, and QA systems.

What carries the argument

The central mechanism is the combination of a clinical gradient-weighted tokenization loss and a two-stage hierarchical fine-tuning schedule. The tokenizer is a VQ-GAN whose reconstruction loss augments pixel-wise $\ell^1$ error with a gradient-matching term $\lambda_{\text{grad}}\|\nabla x_v - \nabla \hat{x}_v\|_2^2$ and a feature-matching term $\lambda_{\text{feat}}\|\varphi(x_v) - \varphi(\hat{x}_v)\|_2^2$ using a pretrained CXR encoder, which is what preserves the fine boundaries and texture that matter for diagnosis. The schedule first optimizes the joint conditional likelihoods $P(x_t \mid x_v)$ and $P(x_v \mid x_t)$ on large paired data, then fine-tunes per task with instruction templates and an $\ell^2$ weight-distance regularizer against catastrophic forgetting. This machinery carries the argument because the paper attributes its performance gains to exactly these components and shows their removal degrades the BLEU score.

What would settle it

Recompute the headline BLEU, ROUGE-L, METEOR, CIDEr, FID, accuracy, and AUROC on the standard MIMIC-CXR held-out split with fixed preprocessing and the baselines' released checkpoints; if MAViLT no longer leads on the majority of metrics, the state-of-the-art claim is falsified.

Watch

Extended reading notes

Core claim

MAViLT is a generative vision-language model built on a pretrained LLM whose embedding table is expanded to include discrete image tokens produced by a VQ-GAN encoder. The paper's central claim is that two additions make this architecture state of the art for chest radiography: a clinical gradient-weighted reconstruction loss that penalizes gradient and feature differences between the original and reconstructed X-ray, and a hierarchical fine-tuning schedule that first pretrains on broad image-text pairs and then refines on task-specific instructions. The loss is $L_{\text{recon}} = \|x_v - \hat{x}_v\|_1 + \lambda_{\text{grad}}\|\nabla x_v - \nabla \hat{x}_v\|_2^2 + \lambda_{\text{feat}}\|\varphi(x_v) - \varphi(\hat{x}_v)\|_2^2$, and the schedule alternates between text generation and image generation objectives. The paper reports BLEU 48.6, ROUGE-L 50.3, METEOR 41.7, CIDEr 2.12 for report generation; FID 21.1 for image synthesis; and 68.5 percent accuracy with 74.6 AUROC for VQA, all above the compared baselines, and ablation results showing each component contributes.

Load-bearing premise

The central claim assumes the reported automatic and human scores were computed on valid held-out splits of MIMIC-CXR and Indiana CXR with preprocessing and baseline settings comparable to MAViLT, yet the paper gives no split definitions, preprocessing, baseline configurations, or evaluation code.

Editorial extensions

If this is right

  • A single MAViLT-style model could replace three separate systems in a radiology workflow: report drafting, image synthesis for illustration or verification, and clinical QA.
  • The clinical gradient-weighted loss is a reusable recipe for medical VQ-GAN tokenizers; other imaging modalities with boundary-critical findings could adopt it directly.
  • Task-adaptive instruction templates imply that new CXR tasks can be added as new prompts without retraining the whole model.
  • If the reported numbers hold, the new comparison points become the reference values for future work on MIMIC-CXR and Indiana CXR.

Reading between the lines

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

  • If the gradient-weighted loss is the main driver of quality, the same term could be ported to CT or MRI tokenizers, where edges and textures carry diagnostic weight.
  • The bidirectional design suggests a consistency test the paper does not run: regenerate an image from a generated report and measure reconstruction FID; low cycle-consistency would expose misalignment between the two directions.
  • The reported efficiency claims, such as no significant increase in computational cost, are asserted without measurements; a deployment decision would need per-task FLOPs or latency numbers.
  • A practical next step would be evaluating on out-of-distribution scanner data or pathologies absent from training, since Indiana CXR is a start but not a full generalization 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 / 5 minor

Summary. The paper proposes MAViLT, a multi-stage adaptive vision-language tuning framework for chest X-ray (CXR) understanding and generation. The framework combines VQ-GAN-based tokenization with a clinical gradient-weighted reconstruction loss, a hierarchical two-stage training strategy, and task-adaptive instruction templates to perform CXR-to-report generation, report-to-CXR generation, and vision question answering (VQA). The authors report state-of-the-art results on MIMIC-CXR and Indiana University CXR across all three tasks, supported by automatic metrics and a human evaluation by radiologists.

Significance. If the reported results were reproducible and correctly benchmarked, a single unified model that generates accurate radiology reports, synthesizes realistic CXRs from text, and answers clinical questions would be a valuable contribution to medical imaging and vision-language research. The proposed loss terms (Eq. 4) and hierarchical training are conceptually reasonable ingredients for such a system. However, the evidence presented in this manuscript is not sufficient to establish the central SOTA claim: the baseline citations are incorrect, the Indiana generalization results are not actually shown in the cited tables, no experimental protocol is specified, and no statistical significance or variance information is provided. The paper therefore does not currently meet the evidentiary standard for its claims.

major comments (5)
  1. [IV-G, Tables I-II] The Indiana generalization results are not actually reported in the tables cited. Section IV-G states that MAViLT achieves a BLEU of 46.7 and an FID of 22.3 on the Indiana dataset, and that 'the results in Tables I and II show' these numbers. However, Tables I and II list BLEU 48.6 and FID 21.1, which are the values used for MIMIC-CXR in Section IV-B. The prose numbers do not appear in any table, so the claim that MAViLT generalizes to an unseen dataset is unsupported by the reported evidence.
  2. [IV-A, references [2],[4],[5]] The baseline comparisons are not verifiable because the cited references do not correspond to the named systems. Reference [2] is a survey on large language models, not UniXGen; reference [5] is a paper on multimodal knowledge graphs, not LLM-CXR; reference [4] is a general investigation of multimodal large language models in medical imaging, not XrayGPT; and RadFM is given no citation at all. Moreover, Section IV-A provides no split definitions, no preprocessing details, no baseline configurations, and no evaluation code. Without correct references and a transparent evaluation protocol, the reported advantages over baselines in Tables I-III cannot be checked.
  3. [Tables I-V] No error bars, confidence intervals, or significance tests are reported for any automatic or human metric. The differences between MAViLT and the strongest baseline in Tables I and III (e.g., BLEU 48.6 vs 46.1; accuracy 68.5 vs 65.1) are small and could plausibly arise from run-to-run variation or different preprocessing choices. The ablation study in Table IV covers only BLEU on the report-generation task, again without variance or significance testing. This does not provide adequate statistical support for the claims of state-of-the-art performance across all tasks.
  4. [III, Eqs. (4), (11), (12)] The method is under-specified to the point of non-reproducibility. The base LLM is never named, the pre-trained CXR encoder φ in Eq. (4) is not identified, the loss weights λ_grad, λ_feat, and λ_reg are not quantified, the Stage 1 pretraining corpus is not described, and the tokenizer training details are absent. These omissions mean that the proposed architecture, as written, cannot be reimplemented and the behavior of the loss function cannot be evaluated.
  5. [IV-D, IV-H, Table V] The human evaluation is described only as average scores by three board-certified radiologists. No details are given about blinding, the number of cases evaluated, the rating rubric, the selection of radiologist raters, or inter-rater agreement. The claim that MAViLT has 'clinical relevance and utility' is therefore not substantiated. Given the central role of this claim in the abstract and conclusion, the absence of basic methodological information is a serious gap.
minor comments (5)
  1. [Throughout] The text contains many spacing and formatting errors (e.g., 'MA ViLT', 'T uning', 'Fr´echet') that should be corrected in a revised version.
  2. [II-B] Reference [18] is cited twice in the same sentence; one citation appears to be a mistake, and the duplicate should be removed or replaced with the intended source.
  3. [IV-C, Table IV] The ablation study reports only BLEU for a single task. Since the paper claims state-of-the-art performance on three tasks, the ablation should be extended to the other tasks or the scope of the claim should be explicitly limited.
  4. [IV-G] The Indiana dataset numbers given in prose (BLEU 46.7, FID 22.3) do not match any table entry. Either a new table should be added with the Indiana results or the prose should be revised to refer to the actual tables.
  5. [References] Several references appear unrelated to the topics they are used for (e.g., reference [11] is a sentiment analysis survey cited for MIMIC-CXR). The reference list needs a thorough revision to match the cited works.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical benchmark comparisons, with no fitted input renamed as a prediction and no equation equivalent by construction to a reported metric.

full rationale

The paper contains no derivation chain in which a predicted quantity is defined in terms of the target metric or fitted to the same data it is later said to predict. The method section defines standard conditional likelihood and reconstruction losses (Eqs. 4-12), and the experimental section reports benchmark scores; none of these equations reduce by construction to the reported BLEU, ROUGE-L, METEOR, CIDEr, FID, accuracy, or AUROC values. The claimed state-of-the-art results are empirical comparisons, and the cited baseline references are not self-citations by the present authors; even if some citations are mismatched or missing (e.g., ref. [2] is a survey rather than UniXGen), that is an evaluation-protocol or attribution defect rather than a circularity. The internal inconsistency between Section IV-G's Indiana results (BLEU 46.7, FID 22.3) and Tables I-II (BLEU 48.6, FID 21.1) undermines reproducibility, but it does not make any claim true by construction. No parameter is fitted to a subset and then used to predict a closely related quantity in a statistically forced way, and no load-bearing premise depends on the authors' own prior work. Therefore the appropriate finding is no significant circularity.

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

No code or configs are released. The core empirical claims depend on unquantified hyperparameters (lambda_grad, lambda_feat, lambda_reg) and an undefined encoder phi. The axioms list the unverified assumptions needed for the SOTA claim to hold.

free parameters (3)
  • lambda_grad
    Introduced in Eq. (4) as the weight on gradient-difference loss; value is never reported or derived.
  • lambda_feat
    Introduced in Eq. (4) as the weight on feature-space loss using an undefined encoder phi; value is never reported.
  • lambda_reg
    Introduced in Eq. (12) to balance catastrophic-forgetting regularization; value is never reported.
assumptions (5)
  • domain assumption Valid held-out splits and preprocessing exist for MIMIC-CXR and Indiana CXR.
    No train, validation, or test split is described in Section IV-A; all evaluation depends on this.
  • ad hoc to paper A pre-trained CXR encoder phi exists and is meaningful.
    Eq. (4) invokes phi(x_v) without defining architecture, training data, or source.
  • domain assumption Stage-1 multimodal pretraining on broad image-text pairs was actually performed.
    Section III-C1 says Stage 1 uses large-scale paired image-text data, but no such corpus is identified.
  • ad hoc to paper Comparison baselines were configured and run fairly.
    Section IV-A2 lists baseline names without implementations, hyperparameters, or citations for several of them.
  • domain assumption Radiologist human evaluation is unbiased.
    Section IV-D gives only average scores; no blinding, sampling, or inter-rater reliability is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Generative Framework for Bidirectional Image-Report Understanding in Chest Radiography." pith.science (2026). https://pith.science/paper/YYMYT7TN

@misc{pith2026250205926,
  author       = {Pith},
  title        = {Pith review of: A Generative Framework for Bidirectional Image-Report Understanding in Chest Radiography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YYMYT7TN}},
  note         = {Machine review of arXiv:2502.05926}
}
read the original abstract

The rapid advancements in large language models (LLMs) have unlocked their potential for multimodal tasks, where text and visual data are processed jointly. However, applying LLMs to medical imaging, particularly for chest X-rays (CXR), poses significant challenges due to the need for precise visual-textual alignment and the preservation of critical diagnostic details. In this paper, we propose Multi-Stage Adaptive Vision-Language Tuning (MAViLT), a novel framework designed to enhance multimodal reasoning and generation for CXR understanding. MAViLT incorporates a clinical gradient-weighted tokenization process and a hierarchical fine-tuning strategy, enabling it to generate accurate radiology reports, synthesize realistic CXRs from text, and answer vision-based clinical questions. We evaluate MAViLT on two benchmark datasets, MIMIC-CXR and Indiana University CXR, achieving state-of-the-art results across all tasks. Human evaluations further validate the clinical relevance and utility of MAViLT, making it a robust tool for real-world medical applications. This work demonstrates the feasibility of leveraging LLMs for multimodal medical imaging while addressing key challenges in vision-language integration.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 16 canonical work pages

  1. [2]

    A survey on large language models: Applications, challenges, limitations, and practical usa ge,

    M. U. Hadi, R. Qureshi, A. Shah, M. Irfan, A. Zafar, M. B. Sh aikh, N. Akhtar, J. Wu, S. Mirjalili et al., “A survey on large language models: Applications, challenges, limitations, and practical usa ge,” Authorea Preprints, 2023

    work page 2023

  2. [5]

    Multimodal reasonin g with multimodal knowledge graph,

    J. Lee, Y . Wang, J. Li, and M. Zhang, “Multimodal reasonin g with multimodal knowledge graph,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Lingui stics (V olume 1: Long Papers), ACL 2024, Bangkok, Thailand, Augus t 11-16, 2024 , L. Ku, A. Martins, and V . Srikumar, Eds. Association for Computational Linguistics, 2024, p...

    work page doi:10.18653/v1/2024.acl-long.579 2024

  3. [4]

    An Early Investigation into the Utility of Multimodal Large Language Models in Medical Imaging

    S. Khan, M. R. Biswas, A. Murad, H. Ali, and Z. Shah, “An ear ly investigation into the utility of multimodal large languag e models in medical imaging,” CoRR, vol. abs/2406.00667, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2406.00667

    work page Pith review arXiv doi:10.48550/arxiv.2406.00667 2024

  4. [1]

    Towards robust ranker for text retrieval,

    Y . Zhou, T. Shen, X. Geng, C. Tao, C. Xu, G. Long, B. Jiao, an d D. Jiang, “Towards robust ranker for text retrieval,” in Findings of the Association for Computational Linguistics: ACL 2023 , 2023, pp. 5387– 5401

    work page 2023

  5. [3]

    Less is m ore: Vision representation compression for efficient video gene ration with large language models,

    Y . Zhou, J. Zhang, G. Chen, J. Shen, and Y . Cheng, “Less is m ore: Vision representation compression for efficient video gene ration with large language models,” 2024

    work page 2024

  6. [6]

    Self - introspective decoding: Alleviating hallucinations for l arge vision- language models,

    F. Huo, W. Xu, Z. Zhang, H. Wang, Z. Chen, and P . Zhao, “Self - introspective decoding: Alleviating hallucinations for l arge vision- language models,” arXiv preprint arXiv:2408.02032 , 2024

    arXiv 2024

  7. [7]

    A tari- gpt: Investigating the capabilities of multimodal large la nguage models as low-level policies for atari games,

    N. R. Waytowich, D. White, M. Sunbeam, and V . G. Goecks, “A tari- gpt: Investigating the capabilities of multimodal large la nguage models as low-level policies for atari games,” arXiv preprint arXiv:2408.15950 , 2024

    arXiv 2024

  8. [8]

    Generative ai for medical imaging: extending the monai fra mework,

    W. H. Pinaya, M. S. Graham, E. Kerfoot, P .-D. Tudosiu, J. D afflon, V . Fernandez, P . Sanchez, J. Wolleb, P . F. Da Costa, A. Patel et al. , “Generative ai for medical imaging: extending the monai fra mework,” arXiv preprint arXiv:2307.15208 , 2023

    arXiv 2023

Show all 27 references
  1. [9]

    Coggpt: Unleashing the power of cognitive dynamics on large language models,

    Y . Lv, H. Pan, Z. Wang, J. Liang, Y . Liu, R. Fu, M. Liu, Z. Wan g, and B. Qin, “Coggpt: Unleashing the power of cognitive dynamics on large language models,” arXiv preprint arXiv:2401.08438 , 2024

    2024 arXiv

  2. [10]

    Echoprime: A multi-video view-informed vis ion- language model for comprehensive echocardiography interp retation,

    M. Vukadinovic, X. Tang, N. Y uan, P . Cheng, D. Li, S. Chen g, B. He, and D. Ouyang, “Echoprime: A multi-video view-informed vis ion- language model for comprehensive echocardiography interp retation,” arXiv preprint arXiv:2410.09704 , 2024

    arXiv

  3. [11]

    Large language models meet text-centric multim odal sentiment analysis: A survey,

    H. Y ang, Y . Zhao, Y . Wu, S. Wang, T. Zheng, H. Zhang, W. Che , and B. Qin, “Large language models meet text-centric multim odal sentiment analysis: A survey,” CoRR, vol. abs/2406.08068, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2406.08068

    2024 doi arXiv

  4. [12]

    Style-aware contrastive learning for multi-style image captioning,

    Y . Zhou and G. Long, “Style-aware contrastive learning for multi-style image captioning,” in Findings of the Association for Computational Linguistics: EACL 2023 , 2023, pp. 2257–2267

    2023

  5. [13]

    Improv ing zero-shot cross-lingual transfer for multilingual questi on answering over knowledge graph,

    Y . Zhou, X. Geng, T. Shen, W. Zhang, and D. Jiang, “Improv ing zero-shot cross-lingual transfer for multilingual questi on answering over knowledge graph,” in Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Ling uistics: H...

    2021

  6. [14]

    Modeling event-pair relations in external knowledge graphs for scri pt reasoning,

    Y . Zhou, X. Geng, T. Shen, J. Pei, W. Zhang, and D. Jiang, “ Modeling event-pair relations in external knowledge graphs for scri pt reasoning,” Findings of the Association for Computational Linguistics : ACL-IJCNLP 2021, 2021

    2021

  7. [15]

    Improving cross-modal alignment f or text- guided image inpainting,

    Y . Zhou and G. Long, “Improving cross-modal alignment f or text- guided image inpainting,” in Proceedings of the 17th Conference of the European Chapter of the Association for Computational L inguistics, 2023, pp. 3445–3456

    2023

  8. [16]

    Visual in-context l earning for large vision-language models,

    Y . Zhou, X. Li, Q. Wang, and J. Shen, “Visual in-context l earning for large vision-language models,” in Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand an d virtual meeting, August 11-16, 2024 . Association for Computational Linguis- ti...

    2024

  9. [17]

    Weak to strong generaliz ation for large language models with multi-capabilities,

    Y . Zhou, J. Shen, and Y . Cheng, “Weak to strong generaliz ation for large language models with multi-capabilities,” in The Thirteenth International Conference on Learning Representations , 2025. [Online]. Available: https://openreview.net/forum?id=N1vYivuSKq

    2025

  10. [18]

    Wolf: Wide-scope large language model framework for CXR understanding,

    S. Kang, D. Kim, J. Kim, H. K. Lee, and S. J. Hwang, “Wolf: Wide-scope large language model framework for CXR understanding,” CoRR, vol. abs/2403.15456, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2403.15456

    2024 doi arXiv

  11. [19]

    LLM-CXR: instru ction- finetuned LLM for CXR image understanding and generation,

    S. Lee, W. J. Kim, J. Chang, and J. C. Y e, “LLM-CXR: instru ction- finetuned LLM for CXR image understanding and generation,” i n The Twelfth International Conference on Learning Represen tations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net, 2024. [Online]. Ava...

    2024

  12. [20]

    Dif f- cxr: Report-to-cxr generation through a disease-knowledg e enhanced diffusion model,

    P . Huang, B. Guo, S. Liang, J. Fu, Y . Wang, and Y . Guo, “Dif f- cxr: Report-to-cxr generation through a disease-knowledg e enhanced diffusion model,” CoRR, vol. abs/2410.20165, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2410.20165

    2024 doi arXiv

  13. [21]

    Advancing text-driven chest x-ray generation with policy-based rein forcement learning,

    W. Han, C. Kim, D. Ju, Y . Shim, and S. J. Hwang, “Advancing text-driven chest x-ray generation with policy-based rein forcement learning,” in Medical Image Computing and Computer Assisted 6 Intervention - MICCAI 2024 - 27th International Conference , Marrakesh, Morocco, Octob...

    2024

  14. [22]

    Addressi ng asynchronicity in clinical multimodal fusion via individu alized chest x- ray generation,

    W. Y ao, C. Liu, K. Yin, W. K. Cheung, and J. Qin, “Addressi ng asynchronicity in clinical multimodal fusion via individu alized chest x- ray generation,” in Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Sys tems 2024,...

    2024

  15. [23]

    Optimising chest x-rays for image analysis by identifying and removing confounding factors,

    S. Aslani, W. Lilaonitkul, V . Gnanananthan, D. Raj, B. R angelov, A. L. Y oung, Y . Hu, P . Taylor, D. C. Alexander, and J. Jacob, “Optimising chest x-rays for image analysis by identifying and removing confounding factors,” in Proceedings of 2022 International Conference on ...

    2022

  16. [24]

    Padchest-gr: A bilingual chest x-ray datas et for grounded radiology report generation,

    D. C. Castro, A. Bustos, S. Bannur, S. L. Hyland, K. Bouzi d, M. T. Wetscherek, M. S´ anchez-V alverde, L. Jaques-P´ erez, L. P´ erez- Rodr´ ıguez, K. Takeda, J. M. Salinas, J. Alvarez-V alle, J. G alant, and A. Pertusa, “Padchest-gr: A bilingual chest x-ray datas et for groun...

    arXiv

  17. [25]

    Gan-based disentanglement learning for chest x-ray rib suppression,

    L. Han, Y . Lyu, C. Peng, and S. K. Zhou, “Gan-based disentanglement learning for chest x-ray rib suppression, ” Medical Image Anal. , vol. 77, p. 102369, 2022. [Online]. Available: https://doi.org/10.1016/j.media.2022.102369

    2022

  18. [26]

    Fine- grained image-text alignment in medical imaging enables ex plainable cyclic image-report generation,

    W. Chen, L. Shen, J. Lin, J. Luo, X. Li, and Y . Y uan, “Fine- grained image-text alignment in medical imaging enables ex plainable cyclic image-report generation,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics ( V olume 1: Long Pap...

    2024 doi

  19. [2024]

    Available: https://doi.org/10.48550/ar Xiv.2411.05085

    [Online]. Available: https://doi.org/10.48550/ar Xiv.2411.05085

    doi arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.