- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes VIGOR, a two-stage reinforcement learning framework for chest X-ray report generation. In the first stage, the model is warmed up with SFT and then trained with GRPO and a verifiable IoU reward to localize abnormal regions. In the second stage, the predicted regions are cropped and fed into the model together with the global image to generate reports. Experiments on multiple public chest X-ray datasets show that VIGOR-4B outperforms open models of similar size and is competitive with closed models. The authors also use ablation studies, a cross-judge evaluation, and a blinded physician study to support the effectiveness of each component. Overall, I think this is a well-motivated and solid paper.
Critical strengths
- The two-stage design, which grounds abnormal regions first and then generates reports conditioned on the cropped regions, is intuitive and well motivated. Feeding localized visual evidence directly into report generation is a reasonable way to strengthen the LLM report generation.
- VIGOR-4B outperforms open models of similar or larger size on most metrics. It also remains competitive with strong closed models such as GPT-5 and Gemini-2.5-Pro, despite having only 4B parameters.
- The cross-judge evaluation and the blinded physician study add credibility to the results.
Critical weaknesses
- The authors should report the length of the generated reports and the number of findings per report, to rule out the possibility that VIGOR's advantage comes from generating less text.
- How are the predicted boxes matched to the ground-truth boxes when their numbers differ? Is there any penalty when the model predicts the wrong number of boxes?
- The paper does not describe the data or the training settings used in the SFT stage.
- VIGOR's Disease score is inconsistent between Table 4 (3.318) and Table 2 (3.138).
Questions for authors
Please see the Weaknesses above.
3 comments
Thanks for the author's effort. I think the rebuttal has addressed my concerns, and I will maintain my score.
We would like to thank reviewer dcZ3 for the positive and constructive comments. We are glad that we resolved the concerns.
W1. The length and number of findings per report should be reported.
A. Thank you for the suggestion. Please see the below table and Appendix Table. 8 of our revised manuscript. VIGOR generates longer reports (46.8 words) with more findings (3.22) than both MedGemma-4B (39.2 words, 2.84 findings) and MedGemma-4B + SFT (42.3 words, 2.92 findings). Importantly, these results address the concern that VIGOR's performance improvements might simply result from generating shorter reports or omitting clinical findings. Instead, VIGOR achieves superior performance while generating more detailed reports containing a greater number of findings. This suggests that its improvements are not merely attributable to reduced report length or fewer reported findings.
Table 4. Average Report Length and Number of Findings per Report
W2. Box matching and penalties when the number of predicted boxes differs from the ground-truth boxes.
A. Thank you for raising this concern. Although we do not explicitly penalize the number of predicted bounding boxes, we focus on spatial alignment between predicted and ground-truth regions through IoU. Thus, missing or inaccurately localized ground-truth regions result in lower IoU rewards. On the test set, the average number of ground-truth bounding boxes was 1.90, while the average number of predicted bounding boxes was 1.79. Nevertheless, it would be an interesting direction for future research to explicitly optimize the number of predicted bounding boxes.
W3. Data and training settings used in the SFT stage.
A. In Stage 1 SFT of VIGOR, we used 5,839 abnormal cases with ground-truth bounding-box annotations from MS-CXR, VinDr-CXR, and NIH Chest X-ray14, randomly split into training, validation, and test sets at an 8:1:1 ratio. In SFT baseline, we used chest X-rays from MIMIC-CXR and Open-I, with their ground-truth FINDINGS sections serving as the target text for report generation. We randomly split the data into training, validation, and test sets at an 8:1:1 ratio. For both SFT experiments, we fine-tuned the model for three epochs by using cross-entropy loss and the AdamW optimizer with a learning rate of 5 × 10⁻⁶.
W4. Typo.
A. Thank you for pointing this out. The correct score is 3.138, as reported in Table 2 of the manuscript.