acceptodds
Under review as a conference paper at ICLR 2027

Visit-order Exchange Equivariance Modeling for Region-Level Longitudinal Report Generation

Abstract

Automated radiology report generation (RRG) has proven its potential to reduce the workload of radiologists. However, most existing RRG methods rely on single-visit data, overlooking the longitudinal changes in the disease. While existing longitudinal methods track disease progression, they typically rely on global features that limit fine-grained regional correspondence, and they are prone to temporal template shortcuts due to visit-order sensitivity. To address this, we propose VEE-RG, a longitudinal radiology report generation framework that generates visit-order stable reports. Specifically, to overcome spatial misalignments and missed detections caused by patient posture variations, we design a paired-region stabilization module to construct reliable visual correspondence. Furthermore, we propose an exchange-equivariant change factorization module that decouples longitudinal differences into symmetric and antisymmetric components, effectively mitigating order sensitivity and temporal shortcuts. Finally, anonymized region-level contrastive learning is introduced to align modalities, preventing shortcut matching via anatomical names and enforcing authentic pathology learning. Extensive experiments on two large public datasets demonstrate that VEE-RG exhibits more stable temporal modeling capabilities and outperforms competitive methods, achieving a 1.7% improvement in CheXbert F1-score on the MIMIC-CXR dataset and a 4.1% improvement in F1-score on the Chest ImaGenome dataset. Our code is available at https://anonymous.4open.science/r/VEE-RG-C1F4.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Longitudinal Radiology Report Generation, Equivariance, Region-Level Modeling, Medical Vision-Language

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QBEqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents VEE-RG, a region-level longitudinal radiology report generation framework that stabilizes anatomical correspondences across visits, factorizes regional changes into exchange-symmetric and exchange-antisymmetric components, and aligns longitudinal visual and textual representations through anonymized contrastive learning; experiments on Chest ImaGenome and MIMIC-CXR report improved clinical efficacy and better stability under visit-order exchange.

Critical strengths

1.The paper addresses a limitation of existing longitudinal report generation methods. Modeling disease evolution at the anatomical-region level is clinically meaningful and potentially more interpretable than relying only on global image representations.

2.The exchange-equivariant formulation is intuitively appealing. Decomposing longitudinal information into a symmetric component that remains unchanged under visit exchange and an antisymmetric component whose sign reverses provides a reasonable inductive bias for separating change existence from change direction.

3.The experimental analysis is relatively comprehensive. The paper evaluates two public datasets and includes structural ablations, auxiliary-task ablations, interventions on the antisymmetric component, region-level change and direction analysis, visit-order exchange experiments, and qualitative attention visualizations. These experiments generally support the contributions of PRS and ECF.

Critical weaknesses

1.The report from the earlier visit is not necessarily a valid reference for the report that should be generated after exchanging the two visits, because it was written before the later examination existed and may compare with another, even earlier study. For example, if the current report states that “pleural effusion has increased,” a correct reversed description should indicate a decrease; however, the original prior report may only state “small pleural effusion” without describing this reversed change. Therefore, I am not fully convinced that the reported Absolute Gap and TRD results directly demonstrate correct temporal-direction reasoning. This is an important evaluation concern that could be addressed by clarifying the construction of reversed references and providing label-based or human evaluation of the predicted change direction.

2.In PRS, a region is considered valid only when it is detected in both visits. If it is missed in either image, the bounding-box prior and ROI feature are disabled for both visits, even when the detection in the other image is correct. I understand that this design preserves symmetry, but it unnecessarily removes useful localized evidence in exactly the difficult cases PRS is intended to handle. Moreover, the paper does not report how frequently the regional boxes and ROI features become unavailable; if this rate is high, the proposed regional comparison would largely fall back to query-conditioned global C5 features, weakening the claimed benefit of explicit regional correspondence. This limitation could be addressed by reporting the failure rate and comparing one-sided masks or confidence-weighted regional features.

3.On MIMIC-CXR, VEE-RG obtains a lower Tempo F1 than PriorRG, while also underperforming it on BLEU-4, METEOR, and ROUGE-L. Its clinical F1 improvement is mainly associated with higher recall, accompanied by lower precision. I therefore find the evidence for improved temporal reporting less convincing than the evidence for increased finding coverage.

Questions for authors

1.The MIMIC-CXR dataset contains AP/PA frontal images as well as lateral-view images, while VEE-RG relies heavily on anatomical bounding boxes and regional correspondence. Since region-grounding models designed for frontal radiographs may be less reliable on lateral views, could the authors clarify whether lateral images were excluded and only AP/PA images were retained? If lateral images were included, how were their bounding boxes and cross-visit regional correspondences handled, and was the effect of view position evaluated? 2.RGRG is an important early baseline for region-level radiology report generation. Why are the RGRG results not reported on the MIMIC-CXR split? Is this due to differences in regional annotations, data splits, or experimental settings?

3 comments
AuthorsOpenReview ↗

Main Questions and Some Weakness

We thank Reviewer QBEq for the careful review and constructive feedback. We also appreciate your recognition of the clinical relevance of modeling disease evolution at the anatomical-region level, the motivation for our exchange-equivariant formulation, and the experimental evidence supporting the contributions of PRS and ECF. Your comments helped us clarify how the original reports are used in the visit-order exchange evaluation and explain what the clinical and temporal metrics capture. We now clarify and address the weaknesses and questions raised by the reviewer.

Questions

Q1: Are lateral-view images excluded?

Thank you for your attention to image views and the applicability of regional localization. The inputs used in this study are all AP or PA frontal chest radiographs; lateral-view images are not included. We further summarized the static report-generation set under the official MIMIC split:

SplitAPPALateral
Train81,83077,6040
Validation6576110
Test1,6307920

The longitudinal training, validation, and test sets contain 108,207, 934, and 2,260 examination pairs, respectively, and the images at both endpoints are all AP or PA. Consequently, no frontal anatomical boxes are applied to lateral-view images, and cross-visit comparisons strictly occur between aligned canonical frontal regions.

Q2: Why are RGRG baseline results missing on the MIMIC split?

Thank you for pointing this out. We have now added the complete evaluation results under this setting:

MetricBLEU-1BLEU-4METEORROUGE-LP-14R-14F1-14
RGRG0.3270.0860.1360.2250.5050.4520.477
MetricP-5R-5F1-5Tempo F1
RGRG0.5000.5970.5440.249

VEE-RG outperforms the baseline model on all metrics. We have added the RGRG results to Tables 1 and 2 of the paper.

Major Weaknesses

W1: Using the original early report, which lacks a description of the reverse comparison, directly as the evaluation reference for the swap experiment may make the AG and TRD metrics difficult to accurately reflect the model's temporal-direction reasoning ability; a label-based or manual direction evaluation should be added.

Thank you for pointing out that the reference context of the early report may differ. After swapping the order of the two examinations, we designate the early examination as the new generation target and directly score against its original report at that time. For example, suppose that the later image indeed shows more pleural effusion than the earlier image. Viewed in the reverse direction, the earlier image shows less pleural effusion than the later image. The phrase “small amount of pleural effusion” in the original early report and the phrase “decreased pleural effusion” generated after the swap can both be valid, because the former describes the state at the early time point, whereas the latter describes the change between the two examinations.

This also clarifies what TRD can answer: it evaluates whether a generated report better matches the correct target time point. While CE solely measures clinical finding overlap, Tempo F1 explicitly captures directional temporal words. The forward-generated “increased” can match “increased” in the later reference, whereas a reasonable swapped-direction expression such as “decreased” has no corresponding word in the early reference when that reference only states “small amount of pleural effusion.” Therefore, the basis for TRD discrimination comes from the content actually recorded in the reference reports. In the Official Comment, we provide further analysis of the TRD evaluation, including additional examples of CE and Tempo F1, to validate the reasonableness of the swap evaluation.

AG evaluates whether the overall scores are similar under the two input orders. On the 1,074 examination pairs shared by the two methods, VEE-RG achieves F1-14 scores of 0.569 and 0.567 in the forward and swapped settings, respectively, whereas PriorRG achieves 0.544 and 0.447. The corresponding AG values are 0.245 and 9.680 percentage points. VEE-RG obtains higher clinical scores under both orders, and its scores are also closer across the two directions.

We further performed a direction-intervention experiment on 1,537 pairs by fixing all inputs and only negating ECF's direction component (OO). Using our lexical rules, we matched generated directions against ground-truth change labels. Among the 2,266 region slots with unambiguous labels, the number of direction-consistent regions decreased from 201 to 35, while the number of direction-opposite regions increased from 26 to 245. This shows that changing the model's internal direction signal systematically changes the directional expression in the generated text.

W2: PRS may lose useful unilateral evidence by completely discarding bilateral local features when one side is missed; frequency statistics and a comparison with a unilateral gating strategy should be provided.

Thank you for your analysis of the use of unilateral local evidence. PRS employs bilateral gating to ensure symmetric feature extraction, as mixing detection-guided local features with query-guided global features would confound true pathological changes with asymmetric reading biases. Therefore, when unilateral detections fail, both visits uniformly fall back to shared-query global attention, ensuring unbiased longitudinal comparisons while preserving essential spatial and identity cues.

To clarify the scope of this design, we added detection-availability statistics under the same definition as the gating comparison over all 2,260 pairs:

Detection statusAll 65,540 region slots3,880 region slots with explicit improvement/worsening
Detected bilaterally76.70%80.85%
Detected on one side only11.20%8.09%
Undetected bilaterally12.11%11.06%
Triggering PRS fallback23.30%19.15%

Thus, 80.85% of explicitly changed regions still utilize bilateral local evidence rather than the fallback. We also evaluated the suggested unilateral-gating variant—where each visit independently uses its own detection flag for local features or query fallback—under identical parameters, supervision, and training budgets:

Test scope and metricBilateral gatingUnilateral gatingRelative improvement (%)Patient-level paired 95% CI for absolute improvement
Natural unilateral-detection group, State macro-F10.56580.5343+5.90[0.69, 5.64]
Natural unilateral-detection group, Direction macro-F10.74560.7090+5.16[−0.03, 7.78]

“Natural unilateral detection” refers to cases in which the actual detection outcome successfully identifies a specific region on only one side, without any artificial masking intervention. Bilateral gating consistently outperforms the unilateral variant, yielding relative improvements of 5.9% in State macro-F1 and 5.2% in Direction macro-F1. We have added the results and analysis to Appendix A.3 of the paper.

AuthorsOpenReview ↗

Weakness 3

W3.1: VEE-RG lags behind PriorRG on Tempo F1 and some NLG metrics.

Thank you for your attention to the evaluation metrics. VEE-RG's lower BLEU-4, METEOR, ROUGE, and Tempo F1 scores than PriorRG may be because PriorRG additionally introduces patient-specific clinical-context text, CC (Indication and Medical History), during both training and inference. These texts naturally contain highly condensed diagnostic intent and medical-history clues, providing a textual advantage for report generation. To verify this point, during inference on the MIMIC test set, we removed the clinical-context (CC) input from PriorRG and retained only the current chest radiograph, the most recent prior chest radiograph (when available), and the acquisition position of each image. The results are as follows:

PriorRG input settingB-1B-4MTRR-LP-14R-14F1-14Tempo F1GFLOPs
VEE-RG (Ours)0.4120.1160.1640.2750.5080.6060.5530.2681447.03
PriorRG w/o CC0.2610.0450.1200.1980.5160.5550.5350.1891877.62
PriorRG w CC0.3950.1670.1850.3240.5760.5010.5360.2822021.21

The results show that PriorRG's NLG and Tempo metrics drop substantially without CC, while its CE metrics show increased Recall and decreased Precision, with overall F1 remaining relatively stable. This indicates that PriorRG's lexical overlap and temporal-state prediction rely heavily on the CC input. Indeed, the original ablation study of PriorRG also showed that this additional text can substantially improve NLG metrics. In contrast, VEE-RG does not rely on additional clinical text and has lower training and inference overhead.

In addition, NLG metrics alone cannot verify whether the longitudinal model has learned a positional shortcut—generating temporal descriptions based on input slots rather than performing a genuine visual comparison of anatomical regions. We therefore introduced AG and TRD, which evaluate visit-order swap stability and target-time-point discrimination ability, respectively. We further conducted a swap evaluation on the 1,074 examination pairs shared by PriorRG and VEE-RG. The results are shown below:

AG ↓BLEU-1BLEU-4F1-14F1-5T-F1
VEE-RG (ours)0.8400.4190.2450.4943.621
PriorRG5.4976.7949.6806.3957.366
Paired-difference 95% CI (pp)[−6.034, −3.360][−7.271, −5.440][−11.770, −5.981][−8.645, −2.367][−7.207, −0.157]
TRD ↑ (%)P-14R-14F1-14F1-5T-F1
VEE-RG (Ours)30.2634.5939.3933.2944.93
PriorRG24.0232.5435.0627.5643.81
Relative improvement (%)+26.0+6.3+12.4+20.8+2.6
95% CI for absolute improvement (pp)[4.14, 8.20][−0.59, 4.53][1.74, 6.85][3.29, 8.07][−1.59, 3.62]

Relative TRD improvements use PriorRG as the baseline. All confidence intervals are patient-level paired bootstrap 95% CIs for the signed differences VEE-RG − PriorRG, expressed in percentage points. VEE-RG's forward and swapped F1-14 scores are 0.569 and 0.567, respectively, while PriorRG's are 0.544 and 0.447; the corresponding AG values are 0.245 and 9.680 percentage points. This shows that VEE-RG not only obtains higher clinical accuracy under both input orders, but also exhibits greater swap stability. At the same time, VEE-RG consistently outperforms PriorRG on both CE-TRD and Tempo-TRD. Taken together, these results show that VEE-RG has lower computational cost than PriorRG and provides more stable visit-order swapping and target-time-point discrimination. We have added the clinical and temporal TRD comparison with PriorRG to Table 15.

W3.2: The clinical F1 gain mainly comes from higher Recall and lower Precision.

Thank you for your attention to the clinical metrics. VEE-RG achieves relative improvements in clinical F1 of 3.2% over PriorRG on MIMIC-CXR and 8.2% over DDGIP on Chest ImaGenome. A more granular analysis shows that, under the MIMIC-CXR split, its CE-5 Recall achieves a relative improvement of 32.8% over PriorRG (0.737 vs. 0.555) and a relative improvement of 17.5% over DDGIP (0.737 vs. 0.627). In actual clinical diagnosis, the potential risk of false negatives is usually much greater than that of false positives. Recall measures the coverage of positive clinical labels in the reference reports. This substantial gain reflects the model's ability to retain more clinical findings and reduce omissions at the label level, which is particularly important in clinical practice.

Reviewer_kZXDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper targets region-level longitudinal radiology report generation (LRRG). Most LRRG methods use global, image-level features and are sensitive to visit order, producing temporal shortcuts instead of true disease progression. The authors propose VEE-RG, comprising: (1) a Paired-Region Stabilization module (PRS) that stabilizes cross-visit regional features (Faster R-CNN detections + learnable spatial attention over the C5 feature map, with a fallback when detection is missing); (2) an Exchange-equivariant Change Factorization module (ECF) that decomposes longitudinal change into symmetric (E) and antisymmetric (O) components, framed honestly as an inductive bias rather than a clinical swap scenario; and (3) anonymized region-level contrastive learning (text replaced by "this anatomical region") with bidirectional InfoNCE to prevent name-based shortcut matching. Reports are generated per region by a frozen-detector, GPT-2 decoder and assembled in a fixed anatomical order. Experiments on MIMIC-CXR and Chest ImaGenome report +1.7% CheXbert F1 on MIMIC-CXR and +4.1% on Chest ImaGenome.

Critical strengths
  1. Thoughtful shortcut-prevention design. Anonymizing anatomical names in the contrastive objective is a neat, well-argued way to block name-based matching and force pathology-level alignment.
  2. Experimental breadth is good: two datasets, 13 baselines, ablations over structure/inputs/losses/antisymmetric component, plus appendix coverage (k-scaling, Top-L, pair statistics).
Critical weaknesses
  1. Worst complexity-to-benefit ratio. ~7 modules + 4 losses + 2-stage training yield only ~1.7–4.1 F1 points overall, while BLEU-4 / ROUGE-L / Tempo F1 fall behind PriorRG; most per-module ablations change metrics by <2 points.
  2. Outdated stack. Faster R-CNN (2016), GPT-2 (2019), BERT (2019), InfoNCE (2020) — no MLLM, no end-to-end training; an incremental recombination of old parts.
  3. Overstated and sloppy. Abstract claims wins while NLG metrics regress (only looking at clinical F1, ignoring NLG's overall lag).
  4. Lack of comparison with modern MLLM methods, such as XrayGPT, LLaVA-Med, etc.
Questions for authors
  1. What is the marginal contribution of each module at matched compute? Report per-module ablations with parameter/FLOPs budget — currently several components change metrics by <1–2 pts, and the full stack only nets ~2–4 pts F1 with NLG losses.
  2. Why Faster R-CNN + GPT-2 + BERT instead of modern MLLM/LLM and ViT-style backbones? Does the claimed improvement survive when the same contrastive/equivariance objectives are placed on current architectures?
6 comments
AuthorsOpenReview ↗

clarity of presentation and some weaknesses

We thank Reviewer kZXD for the careful review and constructive feedback. Your comments helped us clarify the method and its design motivations, quantify the contributions and computational costs of individual components, and strengthen the experimental evaluation with alternative backbones and modern MLLM comparisons. Below, we address the weaknesses and questions raised by the reviewer.

Clarity of presentation

Thank you for your suggestions on readability. We have refined the method descriptions and organization of the manuscript, added key implementation details, and replaced absolute gains with relative improvements. These revisions clarify the overall workflow, the connections between modules, and the motivation behind our design. Revised and newly added material is marked in blue in the manuscript.

Main Weaknesses

W1.1: The overall complexity is high, whereas the total gains are only approximately 1.7–4.1 points.

Thank you for raising the concern about model complexity. We would like to clarify that several components described in the paper, including BERT, the classification supervision, and the image–text contrastive branch, are used only during Stage 1 training and do not participate in inference. The core additional modules used at inference—PRS, ECF, and the EventAdapter—add only approximately 4.58 million parameters. Freezing pretrained components during Stage 2 further limits the additional optimization cost of our two-stage training procedure.

Furthermore, VEE-RG achieves a relative improvement of 3.2% in clinical F1 over PriorRG on MIMIC-CXR and a relative improvement of 8.2% over DDGIP on Chest ImaGenome. A more granular analysis shows that, on the MIMIC-CXR split, its CE-5 Recall achieves a relative improvement of 32.8% over PriorRG and a relative improvement of 17.5% over DDGIP. In clinical diagnosis, false negatives can carry substantially greater risks than false positives. Recall measures coverage of the positive clinical labels in the reference reports. This substantial gain demonstrates the model's ability to preserve more clinical findings and reduce omissions at the label level, which is particularly relevant to clinical reporting.

W1.2: Most per-module ablations change metrics by <2 points.

Thank you for your thoughtful analysis of each module's contribution. We believe that each component should be assessed in relation to its intended task and the corresponding metrics. We have added the number of parameters optimized during Stage 2 to the supplementary table below. The Full, w/o PRS, and w/o PRS+ECF variants have the same number of parameters optimized during Stage 2, yet the Full model achieves better report generation performance, supporting the value of regional correspondence and change modeling. The contribution of longitudinal information is more evident in temporal metrics: adding EVENT increases Tempo F1 from 0.229 to 0.268, a relative improvement of 17.0%. The regional auxiliary tasks provide further evidence (Table 4 in the manuscript): change supervision increases Change F1 from 0.505 to 0.682, a relative improvement of 35.0%, while direction supervision increases Direction F1 from 0.521 to 0.783, a relative improvement of 50.3%. These results show that the components support regional correspondence, change recognition, and directional expression, respectively; their contributions cannot be fully summarized by a single difference in final report F1. We also report inference computation below. The core modules account for only a small proportion of the total GFLOPs. Together with the parameter count in W1.1, these ablations and computational measurements provide concrete evidence of the contribution and cost of our design.

VariantParameters optimized in Stage 2 (M)BLEU-4F1-14Tempo F1
Full107.800.1160.5530.268
w/o PRS107.800.1090.5490.264
w/o PRS+ECF (ROI-MLP)107.800.1080.5450.260
w/o C5104.650.1090.5480.265
w/o EVENT105.050.1100.5390.229
w/o ROI106.750.0980.5170.268
Inference componentGFLOPs
Visual encoding of both visits171.88
Regional report generation1275.15
Total1447.03

Note: Visual encoding of both visits includes detection on both chest radiographs, region selection, PRS/ECF, and visual adaptation.

AuthorsOpenReview ↗

Weakness 1.3

W1.3: VEE-RG lags behind PriorRG on Tempo F1 and some NLG metrics.

Thank you for your attention to the evaluation metrics. VEE-RG scores lower than PriorRG on BLEU-4, METEOR, ROUGE-L, and Tempo F1. One possible reason is that PriorRG additionally uses patient-specific clinical context (CC), comprising Indication and Medical History, during both training and inference. This text contains concise diagnostic intentions and medical-history cues, providing an additional textual advantage for report generation. To examine this, we removed PriorRG's CC input during inference on the MIMIC-CXR test set, retaining only the current radiograph, the most recent prior radiograph when available, and their respective view positions. The results are shown below.

Method / CC settingB-1B-4MTRR-LP-14R-14F1-14Tempo F1GFLOPs
VEE-RG (Ours)0.4120.1160.1640.2750.5080.6060.5530.2681447.03
PriorRG w/o CC0.2610.0450.1200.1980.5160.5550.5350.1891877.62
PriorRG w CC0.3950.1670.1850.3240.5760.5010.5360.2822021.21

Without CC, PriorRG's NLG and Tempo scores decrease substantially. For clinical efficacy, Recall increases while Precision decreases, leaving overall F1 relatively stable. These results show that PriorRG's lexical overlap and temporal-state prediction depend heavily on its CC input. The ablation study in the original PriorRG paper also shows that this additional text substantially improves NLG metrics. In comparison, VEE-RG does not require additional clinical text and has lower training and inference overhead.

Moreover, NLG metrics alone cannot determine whether a longitudinal model has learned positional shortcuts—generating temporal descriptions based on input slots rather than actual visual comparisons of anatomical regions. We therefore introduce Absolute Gap (AG) and Temporal Reference Discrimination (TRD) to evaluate stability under visit-order exchange and discrimination of the intended time point, respectively. We further evaluate PriorRG and VEE-RG on their common subset of 1,074 visit pairs, with the results shown below.

AG ↓ (percentage points)BLEU-1BLEU-4F1-14F1-5T-F1
VEE-RG (Ours)0.8400.4190.2450.4943.621
PriorRG5.4976.7949.6806.3957.366
95% CI for the absolute difference (percentage points)[−6.034, −3.360][−7.271, −5.440][−11.770, −5.981][−8.645, −2.367][−7.207, −0.157]
TRD ↑ (%)P-14R-14F1-14F1-5T-F1
VEE-RG (Ours)30.2634.5939.3933.2944.93
PriorRG24.0232.5435.0627.5643.81
Relative improvement (%)+26.0+6.3+12.4+20.8+2.6
95% CI for the absolute difference (percentage points)[4.14, 8.20][−0.59, 4.53][1.74, 6.85][3.29, 8.07][−1.59, 3.62]

Relative improvements use the corresponding PriorRG score as the denominator. Confidence intervals remain patient-level paired bootstrap 95% CIs for the absolute difference, VEE-RG − PriorRG, expressed in percentage points. VEE-RG achieves forward and swapped F1-14 scores of 0.569 and 0.567, respectively, compared with 0.544 and 0.447 for PriorRG. Their corresponding AG values are 0.245 and 9.680 percentage points, a relative reduction of 97.5%. Thus, VEE-RG achieves higher clinical accuracy in both input orders and stronger stability under visit-order exchange. It also outperforms PriorRG across all reported clinical and temporal TRD metrics. Together, these results show that VEE-RG combines lower computational cost with stronger stability under visit-order exchange and better discrimination of the intended time point. The TRD analysis on the common 1,074 pairs has been added to Table 15 of the paper.

AuthorsOpenReview ↗

weakness 2 4

W2: The technical backbones are dated, the method appears to be an incremental combination of existing components, and training is not end-to-end.

Thank you for your comments on the technical backbones. We follow RGRG's regional detection and generation framework, using established backbones as a common foundation to examine the benefits of regional longitudinal modeling while reducing the confounding effects of simultaneously replacing the detector and language model. Our core contribution is to introduce explicit structural principles for regional longitudinal modeling and to examine their effects through auxiliary supervision, generation interventions, and parameter-matched controls. To our knowledge, VEE-RG is the first method to introduce visit-order exchange equivariance constraints as a structural inductive bias for region-level longitudinal radiology report generation. The methodological contribution lies in the structural constraints developed for the longitudinal task and the experimental connection between these constraints and generation behavior. The exchange constraints operate on paired regional features, and their mathematical definition does not depend on the specific internal mechanisms of Faster R-CNN or GPT-2. This provides a foundation for combining our approach with more advanced visual and language backbones. We further conducted experiments using the more advanced CheXDetector [1] and the larger Llama language model [2]:

MethodB-1B-4METEORROUGE-LP-14R-14F1-14Tempo F1GFLOPs
VEE-RG0.4120.1160.1640.2750.5080.6060.5530.2681447.03
VEE-RG (CheXDetector + GPT-2)0.4000.1080.1550.2600.4790.5610.5170.2722,077.81
VEE-RG (CheXDetector + Llama)0.4000.1100.1540.2610.4930.5240.5080.28017,654.02

VEE-RG achieves the highest scores on the reported NLG and CheXbert metrics with the fewest GFLOPs. Although the Llama-based variant obtains the highest Tempo F1—likely due to its substantially larger parameter capacity—VEE-RG yields the optimal trade-off between performance and computational overhead.

We also thank you for raising this point about end-to-end training. In our formulation, separating representation learning from language generation reflects a deliberate trade-off: it prevents generation objectives from interfering with exchange-equivariant representations and reduces overall optimization cost.

W3: The abstract and overall conclusions overstate the findings.

Thank you for your attention to the accuracy of our claims. The abstract emphasizes clinical F1 because this metric directly reflects our objective of improving the modeling of regional clinical content. As discussed in W1.1 and W1.3, VEE-RG also achieves a relative improvement of 32.8% in CE-5 Recall over PriorRG, alongside improvements in exchange evaluation on the larger common subset. These results provide concrete support for our conclusions. To align the abstract more clearly with the evidence, we have revised the abstract and conclusion to emphasize the advantages in coverage of clinical findings, clinical F1, and discrimination of the intended time point.

W4: VEE-RG lacks comparisons with modern MLLMs.

Thank you for this suggestion. We have added comparisons with modern MLLMs, including XrayGPT, LLaVA-Med, and CXR-LLaVA, using the same test set and evaluation pipeline. The results are shown below.

MethodB-1B-4METEORROUGE-LP-14R-14F1-14Tempo F1GFLOPs
VEE-RG0.4120.1160.1640.2750.5080.6060.5530.2681447.03
XrayGPT [3]0.3310.0510.1280.2250.3510.3270.3390.1562,804.03
LLaVA-Med [4]0.2650.0280.1080.1930.2730.2670.2700.13210,511.49
CXR-LLaVA [5]0.2530.0530.1130.2400.4460.4000.4220.15117,655.51

VEE-RG substantially outperforms these MLLM baselines across all reported performance metrics while using the fewest GFLOPs, demonstrating its advantages for longitudinal report generation.

[1]–[5] These references have been added to the revised paper.

Reviewer_cqPJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VEE-RG, a region-level longitudinal radiology report generation framework designed to improve anatomical correspondence and temporal modeling across consecutive chest X-ray examinations. The framework consists of three main components: (1) a Paired-Region Stabilization (PRS) module that combines anatomical region features with shared spatial queries and a global feature fallback mechanism; (2) an Exchange-Equivariant Change Factorization (ECF) module that decomposes longitudinal representations into symmetric and antisymmetric components; and (3) an anonymized region-level contrastive learning strategy intended to encourage pathology-aware visual-textual alignment. A two-stage training procedure is adopted, followed by region-wise report generation using a GPT-2 decoder. The method is evaluated using MIMIC-CXR and Chest ImaGenome, with experiments covering natural language generation, clinical efficacy, temporal consistency, component ablations, and visit-order exchange analysis. The reported results show improvements in CheXbert F1 and several temporal attribution measures, although performance gains are not consistent across all metrics or baseline comparisons.

Critical strengths
  1. The paper addresses two relevant challenges in longitudinal radiology report generation: establishing correspondence between anatomical regions across examinations and distinguishing directional changes in clinical findings. Both are important for generating clinically meaningful comparative reports.
  2. Beyond conventional generation and clinical metrics, the paper includes auxiliary-task ablations, interventions on the antisymmetric representation, and visit-order exchange experiments. These analyses provide useful initial evidence for the proposed temporal modeling strategy.
Critical weaknesses

Major Weakness

  1. Limited methodological novelty and an insufficient connection between the proposed equivariance principle and the final generation objective. The main methodological contribution appears to be an incremental combination of existing techniques rather than a fundamentally new learning principle. Region-level report generation and anatomical region selection largely follow RGRG, while the PRS module builds on conventional region features, spatial attention, and fallback representations. Similarly, symmetric and antisymmetric feature decomposition can be obtained through standard averaging and differencing operations, as formulated in Eq. (4). Related longitudinal methods, including MARE, have also investigated regional alignment under anatomical and positional variations. More importantly, the theoretical justification for exchange equivariance is limited to intermediate representations. Although the construction ensures that the symmetric component remains unchanged and the antisymmetric component changes sign when visits are exchanged, this property does not necessarily extend to the final generated reports. In Section 3.5, the decoder additionally receives current-image ROI features, global visual features, and region identity embeddings. Furthermore, the EVENT representation undergoes additional transformations before being supplied to an autoregressive decoder. No formal analysis establishes how the complete generation function should behave under visit-order exchange, or why the proposed feature decomposition necessarily prevents the decoder from learning temporal shortcuts. The experimental design does not fully isolate the benefits of exchange equivariance from those of additional temporal features, auxiliary supervision, or model capacity. In particular, comparisons against parameter-matched non-equivariant temporal fusion modules, ordinary signed/absolute feature differences, and visit-swap data augmentation would be important for establishing whether exchange equivariance itself is responsible for the reported improvements. The claims associated with PRS and anonymized contrastive learning also require stronger justification. PRS is presented as addressing missed detections and spatial misalignment, yet the report generator still relies on current-image ROI features, and inference selects detector-valid regions. Thus, it is unclear how the proposed fallback mechanism recovers clinically relevant findings when a region is missed by the detector. No quantitative experiment specifically evaluates robustness under detection failures or varying spatial misalignment. Additionally, the contrastive objective in Eqs. (6)–(8) defines positives according to shared canonical anatomical region IDs. Consequently, samples corresponding to the same anatomical region may be treated as positives despite containing different or contradictory pathological findings. This raises an important conceptual question about whether the objective primarily encourages anatomical identity matching rather than the claimed pathology-aware alignment. Anonymizing anatomical names in the text alone does not establish that anatomical shortcuts have been eliminated. A comparison with conventional region-level contrastive learning without anonymization is also absent. Taken together, these issues weaken the evidence that the proposed components deliver a sufficiently distinctive methodological contribution for ICLR, beyond an application-specific integration of existing architectural and representation-learning techniques.

  2. Several central claims are not adequately supported by the reported experiments, and the evaluation protocol raises concerns. First, the improvements over competitive longitudinal baselines are inconsistent. In Table 1, VEE-RG achieves a CheXbert F1-14 score of 0.553 on MIMIC-CXR, compared with 0.536 for PriorRG. However, its BLEU-4 decreases from 0.167 to 0.116, METEOR from 0.185 to 0.164, and ROUGE-L from 0.324 to 0.275. Table 2 further shows that its Tempo F1 (0.268) is below that of PriorRG (0.282). Although clinical F1 improvements are valuable, these substantial trade-offs make it difficult to conclude that the proposed framework consistently improves longitudinal report generation. The decrease in clinical precision from 0.576 to 0.508 also deserves attention, particularly because additional false-positive findings may be consequential in medical reporting. Second, the visit-order exchange analysis does not conclusively demonstrate improved generalization or reliable temporal attribution. The Absolute Gap metric in Eq. (14) measures the difference between forward and reverse evaluation scores, but a small difference does not necessarily indicate accurate temporal reasoning: a model can perform similarly poorly in both directions and still achieve a small gap. Although Temporal Reference Discrimination (TRD) partially addresses this limitation, the results remain relatively modest. In Appendix Table 7, VEE-RG correctly attributes only 46.12% and 45.21% of forward and reverse comparisons, respectively, while just 13.70% of paired cases are correctly attributed in both directions. These results appear substantially less convincing than the strong conclusions about accurate time-point attribution and highly stable longitudinal reporting. There is also a conceptual distinction between exchanging visit order and evaluating a clinically meaningful reverse temporal task. The original prior report was written in its actual historical context, not as a description of changes relative to a future examination. Therefore, comparing swapped-input generations directly against the original prior report may introduce a mismatch between the intended counterfactual task and its reference. The paper does not sufficiently discuss this limitation. Third, the representativeness of the exchange evaluation is unclear. Appendix Table 9 reports 2,260 longitudinal test pairs, but only 219 pairs belong to the strict intersection used for three-model comparisons. This relatively small common subset raises concerns about selection bias and whether the reported exchange stability generalizes to the full longitudinal population. Confidence intervals or statistical significance analyses are also absent. Finally, the claimed validation on two datasets should be interpreted carefully. Chest ImaGenome is constructed from MIMIC-CXR, and Appendix Table 8 shows identical total numbers of static samples and longitudinal pairs under the two splitting protocols. These experiments therefore do not constitute independent external validation across different institutions or imaging distributions. This is particularly relevant because Section 3.3 suggests that exchange equivariance could improve generalization under real-world distribution shifts, such as noisy prior examinations and inter-hospital variations, without providing corresponding distribution-shift experiments. Overall, the experiments demonstrate certain benefits of the proposed components, but they do not sufficiently establish the broader claims regarding robust temporal reasoning, shortcut mitigation, and generalization. This gap between the claims and the available evidence is a substantial concern for acceptance.

Minor Weaknesses

  1. Insufficient reporting of methodological details. Several implementation choices require clarification for reproducibility. For example, the precise construction and weighting of multi-positive contrastive targets, the definition of valid change/direction labels, and the training or freezing policy of the region selector are not described in sufficient detail.
  2. Limited statistical analysis. The reported improvements are based on point estimates without variance across random seeds, confidence intervals, or significance testing. This is particularly relevant for relatively small gains such as the 0.005 improvement in mean change AUROC reported in Appendix B.2.
  3. Inconsistent attribution of attention mechanisms. Figure 3 describes the visualizations as "ECF attention maps," whereas the spatial attention mechanism is explicitly defined within PRS in Eqs. (1)–(2). The authors should clarify which module produces these attention maps and how they are computed.
  4. Ambiguous percentage reporting. The abstract and Section 4.3 describe the improvement in clinical F1 as 1.7% on MIMIC-CXR and 4.1% on Chest ImaGenome. Based on Tables 1 and 2, these are absolute differences of 1.7 and 4.1 percentage points, respectively, rather than relative percentage improvements. The terminology should be corrected.
  5. Several minor language and presentation issues. In Section 4.5, "VEE-RG yields most balanced gaps" should be revised to "VEE-RG yields the most balanced gaps." In the same section, the statement that models "generate the prior report" after swapping inputs is imprecise; the swapped outputs are evaluated against the prior report, but are not guaranteed to reproduce its content. Some generated examples in Appendix B.4 also contain awkward comparison phrases and redundant clinical descriptions, which deserve additional error analysis.
  6. Limited interpretability of certain supplementary results. Appendix Figure 4 displays independently min–max-normalized performance metrics, making the absolute performance differences across region-selection settings difficult to assess. Providing the corresponding numerical results would make the analysis more informative.
Questions for authors
  1. Clarification and additional evidence regarding equivariance. Could the authors formally explain how the exchange-equivariant properties of the intermediate ECF representations affect the complete report generation function? In particular, what prevents the decoder from ignoring the antisymmetric features or learning positional shortcuts through the current-image features? A parameter-matched comparison against non-equivariant temporal fusion and visit-swap augmentation would materially strengthen the novelty and causal justification of the proposed design.
  2. Clarification of the proposed alignment mechanisms. How does PRS handle cases where an anatomical region is missed in the current examination, given the continued dependence on current ROI features and detector-valid region selection? Additionally, why are all samples sharing a canonical region ID treated as contrastive positives irrespective of pathology? Could the authors provide direct evidence that the proposed anonymization strategy improves pathology-sensitive alignment rather than primarily reinforcing anatomical correspondence?
  3. Clarification and validation of the temporal evaluation. Why does the strict exchange-evaluation subset contain only 219 pairs out of 2,260 available longitudinal test pairs? Could the authors report the characteristics of excluded samples, provide uncertainty estimates for the exchange metrics, and explain the appropriateness of using the original prior report as the reference for reversed inputs? These details are important for assessing the strength and generalizability of the claimed temporal stability.
9 comments
AuthorsOpenReview ↗

Clarification on Methodological Novelty and Exchange-Equivariance

We thank you for your efforts and insightful comments. We address your major concerns below.

Major Weaknesses

W1.1: Clarification on the methodological novelty and distinction from existing longitudinal report generation methods.

Thank you for this suggestion. VEE-RG aims to develop a novel longitudinal report generation method by stably pairing corresponding anatomical regions across prior and current images and incorporating visit-order exchange equivariance as a structural inductive bias. It decomposes the longitudinal features of corresponding regions into symmetric variables (which remain unchanged when the input order is swapped) and antisymmetric variables (which change sign when the input order is swapped), and supervises both representations accordingly. The method achieves encouraging results on clinical and temporal metrics. In addition, the swap evaluation suggests that it reduces cases in which disease changes are inferred from the input order rather than from an actual comparison of corresponding regions. To the best of our knowledge, VEE-RG is the first method to explicitly encode visit-order exchange equivariance as a structural inductive bias in regional change representations for region-level longitudinal radiology report generation.

MARE's ARA also addresses spatial misalignment across visits. However, it does not explicitly establish one-to-one correspondences between the same anatomical regions, and therefore cannot guarantee that its latent features consistently correspond to the same anatomy. In contrast, our explicit inter-region alignment associates longitudinal changes with specific anatomical regions. We also follow RGRG's region-level generation and region-selection pipeline and use the same backbone in the main experiments, allowing us to focus on the benefits of longitudinal regional modeling while reducing confounding factors from changing the backbone at the same time. We further evaluate our method with a more advanced detector and a more advanced LLM, and add these experiments to the paper. See Section 4.6 of the paper.

W1.2: Clarification on the scope of exchange-equivariance guarantees and their implications for report generation and temporal shortcuts.

Thank you for pointing this out. We impose the strict exchange-equivariance constraint on the intermediate regional change representation; we do not claim that the downstream report-generation task is strictly equivariant. Standard report-generation metrics show that VEE-RG produces clinically competitive reports, while the swap evaluation shows more accurate time-point attribution and more stable performance when the input order changes. These experiments demonstrate the generator's strong performance under visit-order swapping and show measurable downstream benefits from the explicit structural constraint on the intermediate representation. Thus, we do not prove exchange equivariance for the generator itself; instead, we evaluate the effect of the representation design through the report-generation task. Nor do we claim that the factorization necessarily prevents the decoder from learning temporal shortcuts. Rather, the swap evaluation shows that the model can reduce reliance on the input order alone to infer disease changes, without actually comparing corresponding image regions. The swap evaluation is presented in Section 4.5 of the paper.

Alongside longitudinal changes EVENT, the decoder receives local ROI, global visual context, and anatomical identity embeddings. To verify if it still utilizes directional representations despite these inputs, we intervene on 1,537 valid pairs during inference. By zeroing or reversing only the antisymmetric component (OO) while preserving the rest of the generation process, we use predefined lexical rules to assess whether the generated text matches or opposes the original direction labels. The results are:

ConditionDirectional expressions consistent with labelsDirectional expressions opposite to labels
Normal (O)20126
(O = 0)7175
(-O)35245

The results show that after setting (O) to zero, the counts of direction-consistent and direction-opposite expressions become similar. After reversing (O), the number of regions whose expressions oppose the original labels increases from 26 to 245, while the number of direction-consistent regions decreases from 201 to 35. This shows that the directional information carried by (O) remains in use by the decoder after the subsequent EVENT transformation.

Note: Directional expressions are identified via automated lexical rules that capture improvement, worsening, and local negations (detailed in Table 12 of the paper's appendix). We apply these same rules in the subsequent anonymization experiment. Here, “direction reversal” denotes a swap between the detected improvement and worsening categories.

AuthorsOpenReview ↗

Control Experiments for Exchange Equivariance and Clarification on PRS (Weakness 1.3 1.4)

W1.3: The experimental design does not fully isolate the benefits of exchange equivariance from additional factors.

Thank you for this comment. We added a parameter-matched non-equivariant temporal-fusion experiment. Specifically, we replaced the exchange constraint with an MLP that fuses the two visits' regional features, their signed difference, and its magnitude, while keeping the other conditions unchanged and matching the parameter count as closely as possible. We call this variant Unconstrained MLP Fusion (UMF). We also added exchange augmentation to UMF: with a 50% probability, we swap the two examinations in a training sample and transform its change label accordingly—unchanged remains unchanged, while improved and worsened are exchanged. We call this variant UMF-Aug. We compare VEE-RG with both controls. The evaluation uses all 1,537 pairs in the MIMIC test set. The results are as follows:

BLEU-1BLEU-4P-14R-14F1-14Tempo F1
VEE-RG0.4070.1080.5310.6230.5730.265
UMF0.4020.1070.5380.6080.5710.248
UMF-Aug0.4030.1070.5410.6020.5700.256
TRDP-14 (%)R-14 (%)F1-14 (%)F1-5 (%)Tempo F1 (%)
VEE-RG31.2035.2340.3434.5644.15
UMF30.0635.0339.7234.3943.01
UMF-Aug30.0334.9139.6933.8043.03
Gain (%)+3.8+0.6+1.6+0.5+2.6
Patient-level 95% CI for VEE-RG − UMF (pp)[0.069, 2.246][-1.228, 1.380][-0.590, 1.990][-0.992, 1.238][-0.407, 2.634]

Compared to the two parameter-matched variants, VEE-RG yields modestly higher BLEU, R-14, and F1-14 scores, alongside a 2.6% relative improvement in Tempo F1 over UMF-Aug. While UMF-Aug achieves higher precision, VEE-RG's gains in recall are clinically crucial, as they reflect reduced omissions of positive findings. Furthermore, VEE-RG outperforms both variant across all TRD metrics, indicating stronger time-point attribution. These results demonstrate that at a comparable parameter count, explicitly encoding exchange-equivariance as a structural constraint provides a more consistent advantage than visit-swap augmentation alone.

W1.4: Clarification on PRS fallback effectiveness and robustness to missed detections.

Thank you for your analysis. PRS fallback enables longitudinal representation learning even with incomplete detections. During generation, region selection still requires a valid current detection; regions missed in the current visit are not independently generated via fallback. If only the prior detection is missing, current-region evidence remains accessible via the static branch. Furthermore, PRS employs bilateral gating to ensure consistent feature-reading conditions across visits. When a detection is valid at only one visit, mixing detection-guided reading with shared-query reading would introduce bias, confounding true pathological changes with methodological inconsistencies. To prevent this, we enforce a shared anatomical query for both visits in such asymmetric cases. To clarify the scope of this design, we have added detection-availability statistics for the full MIMIC-CXR test set:

Detection statusAll region slotsSlots labeled explicitly as improved/worsened
Detected at both visits76.70%80.85%
Detected at only one visit11.20%8.09%
Detected at neither visit12.11%11.06%
PRS fallback triggered23.30%19.15%

Thus, even among regions with explicit changes, 80.85% utilize detection-guided evidence at both visits. Furthermore, we introduced a unilateral-gating control where each PRS branch independently applies its detection flag: the detected side retains full ROI features, while the undetected side reverts to query-based reading. This variant ensures a fair comparison by strictly matching the parameter count, supervision targets, and training budget of the bilateral-gating baseline.

Test subset and metricBilateral gatingUnilateral gatingGain (%) [patient-level paired 95% CI (pp)]
Naturally unilateral detections, State macro-F10.5660.534+5.90 [0.69, 5.64]
Naturally unilateral detections, Direction macro-F10.7460.709+5.16 [−0.03, 7.78]

“Naturally unilateral detections” refers to instances where a region is detected at only one visit, without any artificial masking. In this subset, bilateral gating achieves a State macro-F1 of 0.566, a 5.9% relative improvement over unilateral gating, and a 5.2% relative improvement in Direction macro-F1. See Appendix A.3 of the paper for details.

AuthorsOpenReview ↗

Clarification on Anonymization Strategy and Contrastive Positives (Weaknesses 1.5–1.6)

W1.5: Does the anonymization strategy improve pathology alignment and genuinely reduce name-based shortcuts?

Thank you for your analysis of the anonymization strategy. During training, anonymization replaces explicit anatomical names while retaining descriptions of findings and their changes, with the goal of reducing the opportunity to match image and text using region names alone. To test whether the model attends to change direction, we compare the Stage 1 representations from the anonymized and non-anonymized models. From the longitudinal ImaGenome test pairs, we select regions with valid detections at both visits and an unambiguous direction cue. For each region, we construct two candidate sentences that differ only in the directional terms, which express opposite directions; for example, “increased opacity in the left lower lung” and “decreased opacity in the left lower lung.” The anatomical name and all other content are identical. We then compute the cosine similarity between the regional change visual representation and each candidate sentence's text representation, and select the better-matching sentence. With the images in chronological order, the sentence describing the correct direction should have higher similarity; after swapping the two images, the sentence describing the opposite direction should have higher similarity. The results are as follows:

ModelForward balanced accuracyReverse accuracyBoth-order accuracy
Anonymized71.60%75.07%58.46%
Non-anonymized68.94%74.86%53.72%

Forward balanced accuracy is the average of the recall for the improved and worsened classes. Reverse accuracy measures how often the opposite-direction sentence is selected after swapping the images. Both-order accuracy is the percentage of regions classified correctly under both input orders. The anonymized model's both-order accuracy increases from 53.72% to 58.46%, a relative improvement of 8.8%. Because the anatomical names are identical in both candidate sentences, name matching cannot distinguish between them. This result shows stronger direction-sensitive image-text matching by the anonymized model and supports the motivation that anonymization reduces anatomical-name shortcuts and promotes alignment with pathological changes.

W1.6: Region-ID-based contrastive positives may pull together contradictory pathologies.

Thank you for pointing this out. The statement in the paper that “samples sharing an anatomical region ID are treated as positives” may be inaccurate. In the current implementation, text-target IDs are assigned based on the complete anonymized regional sentence, while each region retains its own visual sample. Visual samples whose processed sentences are identical share a positive text target. Thus, descriptions of different pathologies or opposite change directions are not treated as positives merely because they come from the same anatomical region.

Comments on the paper

1 comment

We thank the AC and all three reviewers for their time and constructive feedback. VEE-RG aims to mitigate models’ overreliance on the input positions assigned to “prior” and “current” images when inferring disease-change direction in longitudinal report generation. To the best of our knowledge, VEE-RG is the first method to explicitly encode visit-order exchange equivariance as a structural inductive bias in regional change representations for region-level longitudinal radiology report generation. We also introduce AG and TRD to evaluate stability under visit-order exchange and target-time-point specificity. In particular, TRD extends conventional single-reference matching to cross-time-point reference discrimination, providing a new perspective for assessing whether longitudinal models effectively distinguish between two examinations. We summarize the main outcomes of the discussion below.

  1. Strengths recognized by the reviewers. cqPJ and QBEq acknowledged the clinical relevance of modeling disease evolution at the anatomical-region level (cqPJ Strength 1; QBEq Strength 1). QBEq recognized the rationale for the exchange-equivariant design (QBEq Strength 2), while kZXD and QBEq acknowledged the breadth of the experimental evaluation (kZXD Strength 2; QBEq Strength 3).

  2. Evidence connecting structural exchange constraints to generation. VEE-RG achieves encouraging results across all five TRD metrics against the approximately parameter-matched UMF and UMF-Aug controls (cqPJ W1.3). Additional interventions on the antisymmetric component show that the decoder uses directional information carried by the intermediate representation (cqPJ W1.2; QBEq W1). cqPJ explicitly acknowledged that these interventions provide evidence that directional information affects the final reports (cqPJ, follow-up comment on November 18).

  3. Additional backbones, baselines, and computational analyses. We added experiments with alternative backbones (kZXD W2, Q2), comparisons with MLLMs (kZXD W4), RGRG results on MIMIC-CXR (QBEq Q2), and parameter-count and GFLOPs analyses (kZXD W1.1–W1.2, Q1). We also examined possible reasons for VEE-RG’s lower Tempo F1 and lower scores on some text metrics than PriorRG, together with the trade-off between clinical precision and recall (cqPJ W2.1; kZXD W1.3; QBEq W3.1–W3.2).

  4. Clarification of the swap evaluation. We clarified that AG and TRD assess exchange stability and target-time-point specificity, respectively, explained the reverse-reference setup (cqPJ W2.2; QBEq W1), and supplemented sample-coverage and uncertainty analyses (cqPJ W2.4–W2.5).

  5. Additional evidence for anonymization and PRS. Supplementary controls and diagnostic experiments support anonymization’s role in improving direction-sensitive image–text matching (cqPJ W1.5) and PRS’s role in stabilizing regional comparisons across visits. We also clarified the scope of the fallback mechanism (cqPJ W1.4; QBEq W2).

  6. Improved methodological detail and presentation. We reconciled the method definitions, corresponding equations, and actual training implementation (cqPJ W1.6 and the authors’ follow-up response on November 18), standardized the reporting of percentage improvements (cqPJ M4), and improved the writing and layout to strengthen clarity and reproducibility (cqPJ M1, M5; kZXD “Clarity of presentation”).

In subsequent feedback, cqPJ acknowledged that the additional experiments helped address several original concerns, while kZXD recognized the strengthened empirical evidence and improved manuscript clarity (cqPJ and kZXD, follow-up comments on November 18). The key experiments and methodological revisions have been incorporated into the revised manuscript, with changes marked in blue. We appreciate the discussion’s contribution to improving this work and respectfully ask the AC to take all relevant aspects into comprehensive consideration in the overall assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.