Abstract

Medical image re-identification (MedReID) supports longitudinal patient linkage but is prone to shortcut learning and often produces decisions that clinicians cannot interpret in terms of named anatomy. We propose Graph-of-Differences (GoD), a framework that grounds identity comparison in explicit anatomical structure. Each image is represented as an anatomy graph with nodes corresponding to named anatomical regions. For an image pair, soft node correspondence is established and differences are computed over matched regions. A graph-level difference alignment objective links these anatomy-matched differences to the global backbone difference, ensuring retrieval signals are anchored in homologous structures rather than arbitrary spatial tokens. Explanations are defined over named graph nodes and evaluated using node insertion and deletion tests, providing verifiable structure-level evidence instead of unstable pixel heatmaps. On internal benchmarks, GoD improves Rank-1 accuracy by +7.1 percentage points on fundus and +3.1 percentage points on chest X-ray over a strong frozen-backbone baseline, with additional gains in zero-shot external transfer demonstrating improved accuracy and generalization through anatomy grounding.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4934_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WasNic_GraphofDifferences_MICCAI2026,
        author = { Wasalathilaka, Nichula AND Das, Abhijit AND Razzak, Imran AND Mahapatra, Dwarikanath},
        title = { { Graph-of-Differences: Anatomy-Structured Difference Alignment for Medical Image Re-identification } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16889},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes Graph-of-Differences (GoD) for medical image re-identification. The method represents each image as an anatomy graph whose nodes correspond to named anatomical regions, establishes soft correspondence between nodes across image pairs, computes differences over matched anatomy, and aligns those graph-level differences with the global backbone difference. The model is built on top of a frozen MaMI encoder, with training limited to the graph encoder, fusion module, and projection head. The paper also proposes node-level explanations over anatomical structures and evaluates them using insertion/deletion-style faithfulness analysis. Experiments are reported on fundus and chest X-ray datasets in both internal and zero-shot external settings, with gains over the frozen-backbone baseline.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    This is a well-motivated paper on an important problem. In medical image re-identification, retrieval performance alone is not enough; it also matters whether the model’s decision can be related to anatomical structures that a human can inspect. The paper tackles that issue directly by enforcing comparison over named anatomy rather than unconstrained token-space differences. That is a sensible and potentially useful design choice. The technical story is coherent. The anatomy graph construction, soft correspondence, and difference-alignment objective fit together naturally. The ablation study is one of the strongest parts of the paper: simply adding a graph branch is not enough, while correspondence plus alignment produces the main gains. That makes the proposed mechanism more convincing. The explanation analysis is also better than what is often seen in similar papers. Rather than relying only on qualitative saliency maps, the paper includes node insertion/deletion-style faithfulness evaluation with confidence intervals. That gives some evidence that the highlighted nodes are actually used by the model. Finally, the external zero-shot results are encouraging. The fact that gains are still observed on CheXpert and Messidor-2 suggests that the method may be learning somewhat more stable anatomical cues rather than only dataset-specific shortcuts.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    My main concern is the paper’s positioning with respect to prior work. The key claim is that GoD improves on token-space comparison by enforcing anatomical correspondence, but the related-work discussion feels incomplete. In particular, there is a directly relevant prior paper, “Graph Feature Embeddings for Patient Re-Identification from Chest X-Ray Images” (SIBGRAPI 2024), which proposes a graph-based approach for chest X-ray patient re-identification. Even if the proposed formulation here is different, this prior work should be cited and discussed explicitly, since it is close in spirit and problem setting. Without that discussion, the novelty claim feels weaker than it should. A second issue is that the evaluation is somewhat narrow relative to the strength of the claims. The experiments compare mainly against a frozen-backbone baseline and internal ablations. That is a clean design for isolating the effect of the proposed graph-based alignment, and I do appreciate that. However, for a strong methodological paper, I would also want to see comparison against stronger practical baselines, especially end-to-end MedReID systems. This is particularly relevant because the backbone here is MaMI-based, and the original MaMI paper reports stronger results after end-to-end tuning. As written, the paper shows that GoD improves a controlled frozen baseline, but not yet that it is the strongest or most competitive approach overall. The paper’s interpretability claim is promising, but in my view slightly overstated. The node-level explanations are clearly more structured than generic pixel saliency maps, and the insertion/deletion analysis is a meaningful step beyond purely qualitative explanation. Still, “clinically auditable” is a strong claim. The paper does not include clinician studies, expert agreement, or any user-grounded evaluation showing that these explanations actually help humans verify or challenge model decisions. I would therefore frame this more cautiously as improved structure-level interpretability with preliminary faithfulness evidence. Another important limitation is parser dependence. The method relies on modality-specific anatomical parsing to build the graphs, but robustness to noisy or missing masks is not evaluated. Since the practical usefulness of the method depends directly on graph quality, this is not a minor omission. A simple robustness experiment with degraded or partial masks would have strengthened the paper considerably. There are also some presentation issues. The manuscript still contains unresolved cross-references such as “Table ??” and “Tables ??–??,” including in places that matter for the experimental section. This makes the submission feel unfinished. Finally, some of the external improvements, especially on fundus, are fairly small. I would have liked to see confidence intervals or significance testing for those results before drawing a strong conclusion about generalization.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    I think this paper has a genuinely interesting core idea, and the ablations are carefully designed. My main suggestion is to sharpen the novelty claim and make the validation story more complete. In particular, the paper would benefit from a clearer discussion of related graph-based MedReID work, comparison to a stronger end-to-end baseline, and at least one robustness experiment for imperfect anatomical parsing. I would also encourage the authors to tone down the “clinical auditability” claim slightly unless they can provide more direct human-centered evidence.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper presents a thoughtful and technically coherent idea: enforcing anatomical correspondence before computing cross-image differences is a sensible way to structure the MedReID problem, and the ablation study supports the importance of the proposed components. The explanation analysis is also stronger than average for this area. However, I do not think the current version makes a complete enough case for acceptance as a methodological contribution. The novelty position is weakened by incomplete engagement with closely related graph-based re-identification work, the empirical comparison is mainly against a frozen baseline rather than stronger practical baselines, parser robustness is not evaluated, and parts of the manuscript remain incomplete at the presentation level. Overall, I see this as a promising paper that would benefit from revision and a stronger validation story.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The main contributions includes the following:

    1.Anatomy-Structured Difference Alignment: A graph-level objective that establishes soft node correspondence across image pairs and aligns these structured differences with the global backbone signal.

    2.Clinically Auditable Explanations: A shift from unstable pixel-level heatmaps to node-level attributions over named structures, which are quantitatively audited via node insertion and deletion tests.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Novel Methodology: The Graph-of-Differences framework introduces a structured way to ground identity comparisons in named anatomical regions rather than arbitrary spatial tokens. This directly addresses shortcut learning, where models rely on scanner artifacts rather than biology.

    2.Verifiable Explainability: Unlike unstable pixel-level heatmaps, GoD provides node-level attributions over named structures. The authors quantitatively validate this using node insertion/deletion tests, proving that the attributed nodes are causally linked to the model’s decisions.

    3.Strong Empirical Performance: The model demonstrates consistent improvements across multiple modalities. Specifically, it improves Rank-1 by +7.1 pp on fundus and +3.1 pp on CXR over a state-of-the-art frozen-backbone baseline.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.Scope Limitations: Currently, the method is limited to 2D images. Extending this to 3D volumetric data (CT/MRI) would require significantly more complex graph correspondence and higher node counts, which the authors acknowledge as future work.

    2.Dependency on Parsers: The framework relies heavily on modality-specific anatomical parsers (like CheXmask). Failure or inaccuracies in these masks can degrade the quality of the anatomy graph, which remains a bottleneck for deployment in new modalities.

    3.Highly Technical: The paper presents a significant barrier to entry, as the core methodology is difficult to understand without high-level technical knowledge in both graph machine learning and medical imaging.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a significant methodological advancement in Medical Image Re-Identification (MedReID) by addressing the critical gap between high-performance black-box models and clinical auditability. It should be accepted due to its strong empirical results, clear technical contribution, and focus on model faithfulness.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The paper makes a clear and well-motivated contribution to medical image re-identification, tackling a key limitation of existing methods by anchoring identity comparison in matched anatomical structures rather than arbitrary features; improving both interpretability and auditability.



Review #3

  • Please describe the contribution of the paper

    The main contribution of this paper is proposing a new method, Graph-of-Differences, that makes medical image re-identification more interpretable by explicitly grounding comparisons in anatomical structures. Instead of comparing images using global or token-level features, the method builds an anatomy graph where each node represents a named region (e.g., lung, optic disc). It then aligns and compares corresponding regions across images and enforces that these structured, node-level differences explain the overall similarity between images.

    This design helps reduce shortcut learning and improves generalization, especially in cross-dataset settings. In addition, the method provides node-level explanations tied to specific anatomical regions, making the model’s decisions more interpretable and potentially more auditable compared to standard pixel-based approaches.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The paper addresses an important issue: lack of interpretability and shortcut learning in medical re-identification. To my perspective, using anatomy to guide comparison is intuitive and meaningful for clinical use.

    2.The paper provides a more structured form of explainability by offering node-level explanations tied to specific anatomical regions. This makes the interpretations more meaningful and clinically relevant compared to traditional pixel-based heatmaps, which are often unstable. In addition, the authors include quantitative evaluation through insertion and deletion tests to support the faithfulness of these explanations.

    4.The result demonstrates gains across multiple datasets and zero-shot settings and Includes ablation studies and explainability evaluation, supporting the effectiveness of the method.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.There are contents missing in the paper. For example: The experiments section says “All hyperparameters are reported in Table ??” and the explainability section refers to “Fig. 4, Tables ??–??.” There are also potential typo in the paper. For example: ““Interna(ODIR)l”. The author needs to get them fixed. Also, parts of the writing need further polishing to improve readability for a general audience. For example, terms like “R@1” should be clearly defined when first introduced.

    2.It lacks detailed information for reproducibility. The paper does not clearly specify batch size, sampling strategy for metric learning, backbone input preprocessing, image resolution, augmentation pipeline, exact identity classification setup, or how external evaluation is done in practice.

    3.The paper mainly compares its method to a weaker baseline where the backbone is frozen, plus a few of its own variants. Since GoD is built on top of MaMI, a stronger evaluation would include better baselines, like a fine-tuned MaMI or other existing methods. Right now, the improvements are shown against a limited setup, so it’s unclear how much gain comes from the method itself versus the weaker baseline.

    4.The method relies heavily on an anatomy parser, which can be a weakness. If the parser makes mistakes, the graph built from it will also be incorrect, and the model may not work well. In other words, the performance of this method depends on how reliable the parser is.

    5.The per-node decomposition {δi} is described as a natural way to audit the model at the node level. However, the paper later states that propagating these attributions to the final embedding z(x) is left for future work, meaning the explanation is not fully complete yet. It needs further explanation.

    6.Table 2 reports the backbone-only CXR internal Rank-1 as 0.8659, but Table 3 reports backbone-only CXR Rank-1 as 0.8723 ± 0.0005.Not sure what causes this inconsistency and the author needs to explain it clearly

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I think the idea is interesting and relevant, but I would lean toward a weak reject at this stage. There are several issues that make the paper feel a bit incomplete. Some parts are clearly missing, and there are typos and writing issues that hurt readability. The reproducibility is also a concern since key details like batch size, preprocessing, and training setup are not clearly described. On the experimental side, the comparison is mostly against a weaker frozen-backbone baseline and its own variants, so it’s hard to tell how much improvement really comes from the method versus the setup. The method also depends a lot on the anatomy parser, which could be a weakness if the parser is not reliable. In addition, the explanation part is not fully complete since the node-level attribution is not connected all the way to the final output. There are also some inconsistencies in the reported numbers that need clarification. Overall, it’s a promising idea, but the paper needs more polishing and stronger validation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #4

  • Please describe the contribution of the paper

    The paper proposes an anatomy-aware framework for medical image re-identification that enforces identity comparison over corresponding anatomical structures rather than arbitrary features. It introduces a graph-of-differences design, where node-level differences between matched anatomical regions are aggregated and aligned with global feature differences, improving interpretability and reducing shortcut learning.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper has several clear strengths. (1) The motivation is well presented and easy to follow. The authors identify an important limitation of existing MedReID methods that identity matching may rely on shortcut features, while the resulting decisions are difficult to verify against named anatomical structures. The proposed solution is also logically connected to this motivation, moving from anatomy graphs to node correspondence, difference alignment, and node-level explanation. (2) The method is interesting and reasonably convincing. Rather than simply adding a stronger backbone or relying on post-hoc heatmaps, the paper introduces an anatomy-structured comparison mechanism, which makes the identity-matching process more interpretable. (3) The experiments are relatively complete. The evaluation covers both fundus and chest X-ray datasets, includes internal and external settings, and provides ablation studies. The insertion/deletion analysis is also useful because it gives quantitative support for the claimed node-level explanations.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The paper could be further improved in several aspects. (1) The clinical feasibility could be strengthened by including clinician evaluation or reader studies to verify whether the proposed node-level explanations are truly useful in practice. (2) The robustness of the method to anatomy parsers or segmentation quality could be further analyzed, as inaccurate masks or landmarks may affect graph construction and downstream performance. (3) The current evaluation is limited to 2D imaging (fundus and chest X-ray), and it would be valuable to explore the applicability of the framework to 3D modalities such as CT, MRI, or PET/CT. (4) The explanation validation could be further strengthened, as insertion/deletion tests, while useful, do not fully demonstrate that the model relies on clinically meaningful anatomical cues rather than dataset-specific correlations.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (6) Strong Accept — must be accepted due to excellence

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I recommend strong acceptance because the paper presents a clear, novel, and well-motivated contribution to medical image re-identification. It addresses an important limitation of existing MedReID methods by grounding identity comparison in matched anatomical structures rather than arbitrary features, making the decision process more interpretable and auditable. The graph-of-differences formulation is technically interesting, and the experiments are relatively complete, covering both fundus and chest X-ray datasets, internal and external settings, ablation studies, and insertion/deletion analysis. Although clinical reader validation and extension to 3D imaging would further strengthen the work, the novelty, clarity, and experimental support justify a strong accept.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    After reading the authors’ rebuttal, I maintain my recommendation to accept the paper.



Author Feedback

We thank reviewers for recognizing GoD’s motivation, the strong ablation (graph alone is inert; correspondence+alignment drive gains), the faithfulness audit, and zero-shot gains.

[R2.1] Prior graph-based MedReID. The SIBGRAPI 2024 paper “Graph Feature Embeddings for Patient Re-Identification from Chest X-Ray Images” uses graph features as a single-image embedding; it does not establish per-pair correspondence between named anatomy, compute differences over matched nodes, or provide faithfulness-audited node-level explanations. GoD’s novelty is the per-pair Graph-of-Differences—a structured difference objective enforcing identity comparison over homologous anatomy. We will cite and contrast in Sec.1.[R2.2 / R4.3] Fine-tuned baseline. The frozen backbone in Tab.2 isolated GoD’s contribution from MaMI tuning. We now also report end-to-end fine-tuned MaMI: Method CXR R1 Fundus R1 Backbone frozen : 0.8659 / 0.6396 Backbone fine-tuned (MaMI) : 0.8841 / 0.6712 GoD (frozen backbone) : 0.9033 / 0.7111 GoD (fine-tuned backbone) : 0.9089 / 0.7253 GoD’s gain is orthogonal to backbone tuning: on fine-tuned MaMI it still adds +2.5 pp CXR/+5.4 pp fundus, and the best absolute result is GoD+fine-tuned. [R2.3] Interpretability framing. We agree “clinically auditable” overstates current evidence. We will revise throughout to “structure-level interpretability with quantitative faithfulness evidence”—what insertion/deletion tests show — and note full auditability requires radiologist user studies (future work). [R2.4/R3.2/R4.4/R5.2] Parser robustness. Soft correspondence (Eq.5) absorbs parser errors via row-softmax down-weighting mismatched/missing nodes. Stress test (CXR R1/Fundus R1): Perturbation CXR Fundus Clean (default) : 0.9033 / 0.7111 10% nodes dropped randomly : 0.8987 / 0.7044 30% nodes dropped : 0.8821 / 0.6852 50% nodes dropped : 0.8543 / 0.6517 5-px mask dilation : 0.8951 / 0.7038 10-px mask dilation : 0.8782 / 0.6859 No graph (frozen backbone) : 0.8659 / 0.6396 GoD stays above the no-graph baseline through 50% node dropout and 10-px mask noise. Added to Sec-4.[R2.5/R4.1] Presentation. We will fix unresolved cross-references (Table??), the “Interna(ODIR)l” typo, and define R@1 at first use. [R2.6] External CIs. Messidor-2 external gain will carry bootstrap 95% CIs: GoD R1 = 0.6579 [0.6491,0.6664] vs. backbone 0.6499 [0.6398,0.6604].

[R4.2] Reproducibility. We will add: batch 64, BatchHard sampler, 224x224 input, ImageNet normalization, label smoothing 0.1, plus the external-eval protocol. [R4.5] Per-node attribution propagation. Per-node delta_i are read from the graph branch where the alignment loss directly operates, so the insertion/deletion audit targets exactly the component the attributions describe—tight by construction. Propagating through the fused embedding z(x) requires learned credit assignment, which we defer rather than approximate: an approximate propagation would conflate graph-branch faithfulness with fusion gating and weaken the current guarantee. [R4.6] Tab.2/Tab.3 inconsistency. 0.8659 (Tab.2) is a single seed; 0.8723+-0.0005 (Tab.3) is the three-seed mean. We will report the seed-averaged value in both.

[R3.1/R5.3] 3D extension. GoD transfers directly to 3D: replace 2D parsers with volumetric segmentors (TotalSegmentator) and the node encoder with a coord-aware GNN for denser graphs and 3D correspondence. We list this as immediate future work. [R3.3] Accessibility. We will add a short intuition paragraph in Sec.2 introducing the graph-of-differences idea before the formal definitions. [R5.1/R5.4] Reader study and explanation depth. Insertion/deletion tests establish faithfulness; clinician studies establish utility. A forced-choice reader study with collaborating clinicians (GoD node-level vs. pixel-saliency baselines on agreement and verification time) planned in post-camera-ready.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    This is promising and well-motivated work with a strong empirical element. All reviewers agree that the work addresses an important limitation in MedReID and appreciate the explainability contribution and relatively comprehensive experimental validation.

    However, some concerns remain regarding dependence on parsers, robustness, baseline comparisons and interpretability claims. There are also some minor presentation issues to rectify.

    If results exist, the fine-tuned model should be reported; otherwise, the authors must justify the frozen backbone. The authors should also demonstrate some evidence of robustness of the parser and review the claims of interpretability. It is also important to clarify the distinction from existing work.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    N/A

  • Please justify your recommendation.

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This submission proposes a graph-based framework for anatomy-based re-identification, R3 and R5 liked the explainability approach. R2 and R4 has concerns on missing fine-tuned baselines, reproducibility details, and dependence on parser. R2 and R4 did not provide feedback on rebuttal. The rebuttal addressed most of the concerns (showing orthogonal performance from requested baselines). For all these reasons, and with respect to the other submissions, the recommendation is towards Acceptance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The rebuttal directly addresses the key concerns by adding fine-tuned MaMI results, parser-robustness stress tests, external confidence intervals, reproducibility details, a clearer distinction from prior graph-based MedReID work, and a more cautious interpretability framing; therefore, despite remaining limitations on clinician validation and 3D extension, I support acceptance.



back to top