Pith. sign in

REVIEW 4 major objections 6 minor 30 references

Multimodal Large Language Models for Medical Report Generation via Customized Prompt Tuning

T0 review · 4 major / 6 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 MRG-LLM generates instance-specific soft prompts from chest X-ray features, reporting state-of-the-art medical report generation on IU X-ray and MIMIC-CXR.

desk verdict Genuinely novel prompt-customization mechanism, but the SOTA claim is undercut by unreproduced baselines and the paper's own ablation numbers are inconsistent. read the letter →

arxiv 2506.15477 v1 pith:4W6J6APY submitted 2025-06-18 cs.CV

Chunlei Li , Jingyang Hou , Yilei Shi , Jingliang Hu , Xiao Xiang Zhu , Lichao Mou This is my paper
classification cs.CV
keywords multimodallargelanguagemodelmedicalreportgenerationprompttuningcustomizedpromptschestX-rayfrozenLLMMIMIC-CXRIU
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

MRG-LLM is a multimodal large language model that writes radiology reports from chest X-rays. Its central claim is that giving the frozen language model soft prompts customized to each individual image—rather than one shared set of prompts—makes the output reports more accurate. The customization is a conditioned affine transformation: a small network reads the image features and outputs scale and shift coefficients that rescale a book of learnable prompt vectors. On IU X-ray and MIMIC-CXR the paper reports state-of-the-art results, with gains of 6.4%, 11.6%, and 8.3% in BLEU-1, BLEU-4, and METEOR on IU X-ray. If the comparison survives scrutiny, the implication is that prompt personalization is a cheap, effective way to improve medical text generation without retraining the LLM.

What carries the argument

The central object is the learnable promptbook P ∈ $R^{{N×D}}$, a matrix of N soft prompt vectors placed before the text tokens, and the central mechanism is the conditioned affine transformation that resizes it. A small MLP parameter network φ takes the visual feature map from a frozen ConvNeXt-Tiny encoder and outputs the scaling and shifting coefficients, which are then applied to the promptbook. This turns a static, task-wide prompt into an instance-specific prompt that depends on the content of the X-ray, via p'_i = γ_i p_i + β_i or P' = γP + β. It is the load-bearing device that lets a frozen LLM receive visual information through its prompt channel rather than through fine-tuning.

What would settle it

Re-run every baseline in Table 1 with its official code under MRG-LLM's tokenizer, decoding settings, and patient-disjoint splits on both datasets, and check whether MRG-LLM still leads. A decisive smaller check: reproduce XPRONET (the one baseline the paper marks as reproduced) under MRG-LLM's evaluation pipeline and see if the 6.4% BLEU-1 gap on IU X-ray persists.

Watch

Extended reading notes

Core claim

The paper's discovery is that a frozen LLM, when conditioned on soft prompts that are customized to the image at hand, generates better medical reports than the same LLM with fixed prompts. Customization is done by a parameter network φ that maps the visual features to coefficients (γ, β), which then transform the prompt vectors via p'_i = γ_i p_i + β_i (prompt-wise) or P' = γP + β (promptbook-wise). The paper reports that prompt-wise customization raises BLEU-1 from 0.395 to 0.416 on MIMIC-CXR compared with no customization, and that the scale coefficient γ is the main driver: at test time, removing γ drops BLEU-1 by 60.3%. The authors interpret this as evidence that instance-specific prompts let the LLM focus on the pathological findings present in each radiograph.

Load-bearing premise

The state-of-the-art claim rests on the assumption that all baseline scores in Table 1 were produced under the same evaluation protocol as MRG-LLM; only XPRONET is marked as reproduced, so if the other baselines used different tokenization, decoding, or splits, part of the margin may be protocol artifact.

Editorial extensions

If this is right

  • Prompt customization provides a parameter-efficient way to improve medical report generation: only the projection layer, promptbook, and tiny parameter network are learned, while the vision encoder and LLM stay frozen.
  • The two-parameter promptbook-wise variant shows that most of the benefit can be captured by a single global rescaling of the prompt embeddings, suggesting the visual information enters mainly through the scale of prompt vectors.
  • The framework can be transferred to other report domains or imaging modalities by relearning the small learned components without touching the LLM.
  • The ablation results identify the scale parameter γ as the critical channel, implying that future designs should focus on how strongly visual features modulate prompt magnitude.

Reading between the lines

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

  • Because the paper reproduces only XPRONET among the 27 baselines, a strict comparison under identical tokenization and decoding could narrow the reported margins; this is an open question the paper does not resolve.
  • The mechanism is a lightweight form of visual grounding and could generalize beyond medicine to any frozen-decoder vision-language task, such as captioning or visual question answering, where a single model must adapt to diverse inputs.
  • One could test whether the customized prompts act as a learned visual summary by probing them with a classifier that predicts pathology labels from the prompt vectors alone, or by visualizing which base prompts get up- or down-scaled for images with specific findings.
  • A clinical usability study—radiologists blind-rating MRG-LLM outputs against baselines for correctness and editing time—would determine whether the metric gains translate into practice, since BLEU and METEOR do not measure clinical factuality.
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 / 6 minor

Summary. The paper proposes MRG-LLM, a multimodal large language model for chest X-ray report generation. It combines a frozen LLaMA backbone with a ConvNeXt-based visual encoder and a linear projection, and introduces instance-specific soft prompts generated by affine transformations (gamma and beta) predicted from visual features. Two instantiations are presented: prompt-wise and promptbook-wise customization. Experiments on IU X-ray and MIMIC-CXR report state-of-the-art results across BLEU, ROUGE-L, and METEOR, together with ablations on the customization mechanism, the gamma/beta decomposition, parameter-network depth, and the number of learnable prompts.

Significance. If the reported results are reproducible and the baseline comparisons are controlled, the contribution is a simple and parameter-efficient idea: customize soft prompts per input image through a low-dimensional affine modulation, with a two-parameter promptbook variant that is particularly economical. The paper includes a sensible set of ablations, including prompt count and parameter-network depth, and the architecture is clearly described. However, the load-bearing state-of-the-art claim is weakened by the lack of protocol-matched baseline reproduction, absence of error bars or statistical tests, and internal inconsistencies in the ablation reporting; the significance is therefore conditional on these issues being resolved.

major comments (4)
  1. [§3.3, Table 1] The state-of-the-art claim is not sufficiently supported because only XPRONET is marked as reproduced under the authors' evaluation protocol. The manuscript does not state which report section is generated, which tokenizer is used for metric computation, or whether the cited baselines used the same patient-disjoint split for IU X-ray and the same official test split for MIMIC-CXR. On MIMIC-CXR the reported margins are small (BLEU-4 is +0.007 over HERGen and METEOR ties AdaMatch-Cyclic at 0.163), so protocol differences could change the ranking. Please reproduce the key baselines under the authors' protocol, or report the exact settings of each cited baseline and soften the SOTA claim accordingly.
  2. [§3.4, Table 3] The ablation percentages in the text do not match Table 3. From Table 3, training with only beta (gamma removed) drops BLEU-1 from 0.416 to 0.398, a 4.3% relative drop, while training with only gamma (beta removed) drops to 0.409, a 1.7% drop. The corresponding test-time drops are 3.4% (gamma only) and 42.3% (beta only). The stated values of 2.9%, 0.1%, 4.2%, and 60.3% are therefore not reproducible from the table, and the source of the discrepancy should be clarified.
  3. [§3.2, §3.3] No error bars, multiple seeds, or statistical significance tests are reported. Given that the MIMIC-CXR improvements over the best prior baselines are within 0.01–0.02 for several metrics, the authors should provide repeated-run statistics and a paired significance test before claiming state-of-the-art performance.
  4. [§3.4, Table 2] The text reports a CIDEr improvement from 0.224 to 0.258, but CIDEr is not defined in Section 3.2 and no CIDEr column appears in Tables 1–4. Either add CIDEr to the evaluation tables or remove the CIDEr claims.
minor comments (6)
  1. [§3.3] The phrase 'absolute improvements of 6.4%, 11.6%, and 8.3%' is imprecise; these are relative improvements (e.g., BLEU-1 0.529 vs 0.497 is a 6.4% relative gain), not absolute differences. Please rephrase as relative improvements or report the absolute deltas.
  2. [§3.2] The abstract says 'learnable visual encoder,' but Section 3.2 merely says features are extracted with a ConvNeXt-Tiny backbone pretrained on ImageNet-1K; please state explicitly whether the backbone is fine-tuned or frozen during training.
  3. [§3.4] Tables 2–4 and Figure 3 do not state which dataset they use; the numbers in Table 2 match the MIMIC-CXR results in Table 1, but this should be stated in the captions or text.
  4. [Fig. 3] Figure 3 lacks a legend; it is unclear which curve corresponds to BLEU-4 and which to METEOR.
  5. [§3.1, Table 1] In Table 1, METEOR for the proposed method on MIMIC-CXR equals AdaMatch-Cyclic (0.163), which contradicts the statement in Section 3.3 that MRG-LLM outperforms previous approaches across all metrics; please qualify the claim.
  6. [Throughout] There are formatting typos such as 'T able 1' in Section 3.3; please proofread the camera-ready version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the prompt-customization model and SOTA claims are evaluated on held-out data against external baselines, with zero self-citations; the flagged issues are internal-consistency and protocol-comparability concerns, not definitional circularity.

full rationale

The derivation chain is not circular at any load-bearing step. The model computes visual features X, produces affine prompt-modulation parameters (γ,β)=φ(X) via Eqs. (3)–(5), applies p'_i = γ_i p_i + β_i (Eq. 4), and trains with standard cross-entropy (Eq. 7) on the training split; the headline results in Table 1 are evaluated on the held-out IU X-ray (7:1:2 patient-disjoint) and official MIMIC-CXR test splits. No output quantity is defined in terms of a model input, and no fitted parameter is renamed as a prediction: the hyperparameters (50 prompts, 2-layer MLP) were selected by validation ablations, which is standard practice rather than a forced result. The reference list contains no self-citations by Li, Hou, Shi, Hu, Zhu, or Mou, so no self-citation chain or imported uniqueness theorem is load-bearing. The claimed improvements (6.4%, 11.6%, 8.3% on IU X-ray) are arithmetic differences against externally published baseline numbers, not consequences of the method's own definitions. Two concerns are flagged but are not circularity. First, Section 3.4's text states training drops of 2.9% and 0.1% and test-time drops of 4.2% and 60.3%, whereas Table 3 implies 1.7%/4.3% (training) and 3.4%/42.3% (inference), and the text concludes γ is more important while the table indicates the opposite; this is an internal-consistency problem in the ablation reporting, not a reduction of any claim to its inputs. Second, only XPRONET is marked as reproduced among 27 baselines, so tokenizer/split/decode protocol differences could affect the SOTA claim; this is an external-validity and verification concern, not circularity. The central contribution — instance-specific prompt customization via learned affine transformations on visual features — is empirically supported by held-out comparisons and does not presuppose its own conclusions.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on pretrained models, standard benchmarks, and the chosen hyperparameters; no new entities are introduced.

free parameters (2)
  • Number of learnable prompts = 50
    Chosen by hand after ablating values {1, 10, 50, 100, 150}; optimal near the number of visual tokens (49).
  • Parameter network depth = 2-layer MLP
    Selected via ablation; linear and 3-layer MLP were worse.
assumptions (3)
  • domain assumption The frozen LLM (LLaMA) and ImageNet-pretrained ConvNeXt provide useful representations for medical images and text.
    The whole method depends on transfer from general-domain pretraining to medical report generation.
  • domain assumption The evaluation metrics BLEU, ROUGE-L, and METEOR are meaningful measures of report quality.
    These n-gram overlap metrics are used throughout the literature but are known to correlate imperfectly with clinical correctness.
  • domain assumption The train/validation/test splits used are patient-disjoint and match prior work.
    The paper states the splits but does not verify that all comparisons used identical splits.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Large Language Models for Medical Report Generation via Customized Prompt Tuning." pith.science (2026). https://pith.science/paper/4W6J6APY

@misc{pith2026250615477,
  author       = {Pith},
  title        = {Pith review of: Multimodal Large Language Models for Medical Report Generation via Customized Prompt Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4W6J6APY}},
  note         = {Machine review of arXiv:2506.15477}
}
read the original abstract

Medical report generation from imaging data remains a challenging task in clinical practice. While large language models (LLMs) show great promise in addressing this challenge, their effective integration with medical imaging data still deserves in-depth exploration. In this paper, we present MRG-LLM, a novel multimodal large language model (MLLM) that combines a frozen LLM with a learnable visual encoder and introduces a dynamic prompt customization mechanism. Our key innovation lies in generating instance-specific prompts tailored to individual medical images through conditional affine transformations derived from visual features. We propose two implementations: prompt-wise and promptbook-wise customization, enabling precise and targeted report generation. Extensive experiments on IU X-ray and MIMIC-CXR datasets demonstrate that MRG-LLM achieves state-of-the-art performance in medical report generation. Our code will be made publicly available.

Figures

Figures reproduced from arXiv: 2506.15477 by the authors.

Figure 1
Figure 1. Overview of our MRG-LLM framework for medical report generation. 2.1 Model Architecture As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization of a sample with reports from the MIMIC-CXR dataset. The baseline employs learnable prompts without customization, while our two proposed prompt customization approaches demonstrate improved performance. Text with the same color highlights descriptions of the same disease across different methods. generation models, including recent works like HERGen [29] and MedRAT [30] [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 3
Figure 3. Ablation study on the impact of varying the number of learnable prompts. Number of Learnable Prompts We investigate how the number of learnable prompts affects MRG-LLM’s performance on the MIMIC-CXR dataset (cf [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 23 canonical work pages

  1. [1]

    In: Advances in Neural Information Processing Systems, pp

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakan- tan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S.: Language models are few-shot learners. In: Advances in Neural Information Processing Systems, pp. 1877–1901. (2020)

    work page 2020

  2. [2]

    arXiv preprint arXiv:2302.13971 (2023)

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    arXiv 2023

  3. [3]

    Journal of the American Medical Informatics Associ- ation23, 304–310 (2016)

    Demner-Fushman, D., Kohli, M., Rosenman, M., Shooshan, S., Rodriguez, L., An- tani, S., Thoma, G., McDonald, C.: Preparing a collection of radiology examinations for distribution and retrieval. Journal of the American Medical Informatics Associ- ation23, 304–310 (2016)

    work page 2016

  4. [4]

    arXiv preprint arXiv:1901.07042 (2019)

    Johnson, A., Pollard, T., Greenbaum, N.R., Lungren, M., Deng, C., Peng, Y., Lu, Z., Mark, R., Berkowitz, S., Horng, S.: MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs. arXiv preprint arXiv:1901.07042 (2019)

    arXiv 2019

  5. [5]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and Tell: A neural image caption generator. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3156–3164. (2015)

    work page 2015

  6. [6]

    In: NIPS, pp

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NIPS, pp. 6000–6010. (2017)

    work page 2017

  7. [7]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Rennie, S., Marcheret, E., Mroueh, Y., Ross, J., Goel, V.: Self-critical sequence training for image captioning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7008–7024. (2017)

    work page 2017

  8. [8]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Lu, J., Xiong, C., Parikh, D., Socher, R.: Knowing when to look: Adaptive attention via a visual sentinel for image captioning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 375–383. (2017)

    work page 2017

Show all 30 references
  1. [9]

    In: IEEE Conference on Computer Vision and Pattern Recognition, pp

    Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question an- swering. In: IEEE Conference on Computer Vision and Pattern Recognition, pp. 6077–6086. (2018)

    2018

  2. [10]

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

    Cornia, M., Stefanini, M., Baraldi, L., Cucchiara, R.: Meshed-memory Transformer for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10578–10587. (2020)

    2020

  3. [11]

    In: Empirical Methods in Natural Language Processing, pp

    Chen, Z., Song, Y., Chang, T., Wan, X.: Generating radiology reports via memory- driven Transformer. In: Empirical Methods in Natural Language Processing, pp. 1439–1449. (2020) 10 C. Li et al

    2020

  4. [12]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, pp

    Zhang, Y., Wang, X., Xu, Z., Yu, Q., Yuille, A., Xu, D.: When radiology report generation meets knowledge graph. In: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 12910–12917. (2020)

    2020

  5. [13]

    S., Fan, W., Zou, Y.: Exploring and distilling posterior and prior knowledge for radiology report generation

    Liu, F., Wu, X., Ge. S., Fan, W., Zou, Y.: Exploring and distilling posterior and prior knowledge for radiology report generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13753–13762. (2021)

    2021

  6. [14]

    arXiv preprint arXiv:2106.06965 (2021)

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

    2021

  7. [15]

    In: Proceedings of the 59th Annual Meeting of the Asso- ciation for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pp

    Liu, F., Ge, S., Wu, X.: Competence-based multimodal curriculum learning for medical report generation. In: Proceedings of the 59th Annual Meeting of the Asso- ciation for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pp....

    2021

  8. [16]

    In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pp

    Chen, Z., Shen, Y., Song, Y., Wan, X.: Cross-modal memory networks for radiology report generation. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pp. 5904–...

    2021

  9. [17]

    In: Medical Image Computing and Computer Assisted Intervention, pp

    You, D., Liu, F., Ge, S., Xie, X., Zhang, J., Wu, X.: AlignTransformer: Hierarchi- cal alignment of visual regions and disease tags for medical report generation. In: Medical Image Computing and Computer Assisted Intervention, pp. 72–82. (2021)

    2021

  10. [18]

    arXiv preprint arXiv:2102.09777 (2021)

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

    2021 arXiv

  11. [19]

    In: Findings of the Association for Computational Linguistics, pp

    Qin,H., Song, Y.: Reinforced cross-modal alignment for radiology report genera- tion. In: Findings of the Association for Computational Linguistics, pp. 448–458. (2022)

    2022

  12. [20]

    In: European Conference on Computer Vision, pp

    Wang, J., Bhalerao, A., He, Y.: Cross-modal prototype driven network for radiol- ogy report generation. In: European Conference on Computer Vision, pp. 563–579. (2022)

    2022

  13. [21]

    In: Medical Image Computing and Computer Assisted Intervention, pp

    Wang, H., Tang, M., Wang, L., Li, X., Zhou, L.: A medical semantic-assisted Transformer for radiographic report generation. In: Medical Image Computing and Computer Assisted Intervention, pp. 655–664. (2022)

    2022

  14. [22]

    Medical Image Analysis 86, 102798 (2023)

    Yang, S., Wu, X., Ge, S., Zheng, Z., Zhou, S., Xiao, L.: Radiology report generation with a learned knowledge base and multi-modal alignment. Medical Image Analysis 86, 102798 (2023)

    2023

  15. [23]

    IEEE Transactions on Multimedia26, 904–915 (2023)

    Zhang, K., Jiang,H., Zhang, J., Huang, Q., Fan, J., Yu, J., Han, W.: Semi- supervised medical report generation via graph-guided hybrid feature consistency. IEEE Transactions on Multimedia26, 904–915 (2023)

    2023

  16. [24]

    arXiv preprint arXiv:2309.09812 (2023)

    Wang, Z., Liu, L., Wang, L., Zhou, L.: R2GenGPT: Radiology report generation with frozen LLMs. arXiv preprint arXiv:2309.09812 (2023)

    2023 arXiv

  17. [25]

    arXiv preprint arXiv:2303.09117 (2023)

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

    2023 arXiv

  18. [26]

    arXiv preprint arXiv:2311.18681 (2023)

    Pellegrini, C., Özsoy, E., Busam, B., Navab, N., Keicher, M.: RaDialog: A large vision-language model for radiology report generation and conversational assistance. arXiv preprint arXiv:2311.18681 (2023)

    2023 arXiv

  19. [27]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, pp

    Jin, H., Che, H., Lin, Y., Chen, H.: PromptMRG: Diagnosis-driven prompts for medical report generation. In: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 2607–2615. (2024)

    2024

  20. [28]

    In: Association for Computational Linguistics, pp

    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: Association for Computational Linguistics, pp. 9494–9509. (2024) Title Suppressed Due to Excessive Length 11

    2024

  21. [29]

    In: European Conference on Computer Vision, pp

    Wang, F., Du, S., Yu, L.: HERGen: Elevating radiology report generation with Llongitudinal data. In: European Conference on Computer Vision, pp. 183–200. (2024)

    2024

  22. [30]

    In: European Conference on Computer Vision, pp

    Hirsch, E., Dawidowicz, G., Tal, A.: MedRAT: Unpaired medical report generation via auxiliary tasks. In: European Conference on Computer Vision, pp. 18–35. (2024)

    2024

Pith tools

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