acceptodds
Under review as a conference paper at ICLR 2027

GC-NCE: Graph-Constrained Nuisance-Aware Clinical Evidence Bottleneck for Medical Image Classification under Domain Shift

Abstract

Concept bottleneck models (CBMs) predict medical image diagnoses through human-interpretable concepts, but they can remain unreliable under distribution shift: an unconstrained diagnostic head can weight concepts by source-domain correlations rather than their clinical roles, and predicted concept scores can encode nuisance factors that classifiers exploit as shortcuts. We introduce GC-NCE, a framework that constructs clinical evidence graphs from retrieved medical literature and uses them to constrain diagnostic inference. The graphs assign typed roles to concepts: direct findings and target mechanisms provide diagnostic support, differential findings provide inhibition, anatomy modulates evidence visibility, and nuisance influence edges restrict where acquisition effects may be corrected. Four LLM agents build each graph through compiler-gated proposals and cross-agent review, and an observability-aware loop refines concepts that cannot be reliably predicted from images. GC-NCE translates these roles into probabilistic inference over latent disease mechanisms, with adaptive nuisance correction enabled per dataset only when source validation confirms its benefit. On ten domain-shift benchmarks, GC-NCE achieves the highest mean out-of-distribution AUROC in both chest X-ray (78.5%) and skin lesion (84.5%) modalities.

Primary area:
interpretability and explainable AI
Keywords:
Concept Bottleneck Models, Interpretability, Knowledge Graphs, Domain Generalization, Shortcut Learning, Medical Image Analysis

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_aqKZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies generalization in medical image classification when associations between disease and demographic or acquisition attributes change. The authors construct an evidence graph from PubMed literature, assigning types and clinical roles to its nodes and edges. An LLM extracts weak concept labels from reports, and logistic-regression concept readers are trained on frozen medical image embeddings. During diagnostic training, the image encoder, concept readers, and graph structure remain fixed. A graph-constrained probabilistic model combines supporting findings, competing diagnoses, latent mechanisms, and anatomical visibility, while restricting nuisance correction to permitted edges. Whether to apply correction is selected using worst-group accuracy on source-domain validation data. On ten attribute-correlation reversal tasks adopted from KnoBo, the method achieves mean OOD AUROCs of 78.5% for chest X-rays and 84.5% for skin lesions.

Critical strengths
  1. The method assigns different diagnostic roles to various clinical relationships, including supporting evidence, competing explanations, anatomical visibility, and the influence of confounding factors. This represents a reasonable alternative to classifiers that arbitrarily exploit correlations among conceptual scores. The decision to enable correction is based on validation results from the source domain, and the potential loss of useful information due to correction is also taken into account.

  2. The experiments cover two imaging modalities and ten existing distribution-shift tasks. Comparisons with conceptual models, linear classification heads using the same reader, graph relation role perturbations, and various correction strategies provide evidence for the effectiveness of structural constraints.

Critical weaknesses
  1. Some important details regarding the construction of evidence graphs remain unclear. The paper specifies retrieval budgets and tool protocols but does not fully explain how initial evidence and queries are selected, how many results are returned each time, how they are sorted and filtered, or how to determine whether evidence coverage is sufficient. The negotiation process may also retain the original content when disagreements remain unresolved. Structural and citation checks do not prove that these disputed relationships are clinically valid, and how they contribute to the final diagnostic calculation remains to be clarified.

  2. The handling of conflicts and disagreements in concept annotation requires clarification. The paper retains the original content when disagreements during graph construction remain unresolved, but it is unclear how these disputed relationships affect subsequent concept annotation and diagnosis. Furthermore, the paper does not sufficiently justify the choice of AUROC ≥ 0.65 as the criterion for retaining concept readers or examine how varying this threshold affects the number of retained concepts and final diagnostic performance.

Questions for authors
  1. How are the initial evidence and search queries chosen for each diagnosis? Please clarify the number of results returned per search, the ranking and filtering rules, and the stopping criterion. The appendix limits the number of searches per agent call, but how is coverage of relevant findings and differential diagnoses assessed? Search records linking the retrieved evidence to the final graph relationships would help clarify this process.

  2. When a disagreement remains unresolved, do the disputed nodes or edges still participate in concept labeling and diagnostic inference? Why are the original relationships retained, and does removing them affect performance? For concept labels, have you assessed false negatives caused by treating unmentioned or uncertain findings as absent? Please also explain how contradictory reports or inconsistent LLM labels are handled, and whether these decisions have been checked against human annotations or independent concept labels.

  3. Why was AUROC ≥ 0.65 chosen as the criterion for retaining a concept reader? Please explain how this threshold was determined and whether alternative thresholds were evaluated. A sensitivity analysis showing how the threshold affects the number of retained concepts and final OOD performance would help justify this choice.

  4. What temperatures and sampling settings are used by the LLMs during retrieval and graph construction? When the full process is repeated for the same disease, how much do the retrieved literature and resulting nodes and edges change? Were multiple independent retrieval and graph-construction runs performed, and were the mean and standard deviation of the resulting OOD performance reported?

6 comments
AuthorsOpenReview ↗

Response to Weakness 1(i) and Question 1(i)

Weakness 1(i) & Question 1(i): Retrieval and graph-construction details are unclear.

Thank you for the constructive comment. As suggested, we describe how a graph is built from start to end, and we summarize the retrieval settings in Table R6.

The corpus and the search tool follow KnoBo (NeurIPS-24) [1]. The corpus consists of articles from the PubMed Central Open Access subset, split at sentence boundaries into passages of about 1,000 characters. It is searched with BM25 as in MedRAG (ACL-24 Findings) [2]. The graph of one diagnosis is built in six steps.

  1. The system issues a fixed list of template queries for the diagnosis (11 for chest X-ray, 8 for skin). They ask about the pathology, anatomy, findings and differential diagnoses of the disease, about image acquisition, and about age, sex and race. An example for Atelectasis is "chest x ray Atelectasis pathology physiology anatomy radiographic findings differential diagnosis". The returned passages are the initial evidence.
  2. The Graph Initialization Agent reads the initial evidence and proposes the mechanisms of the disease, the findings that support the diagnosis, and the anatomy in which each finding is seen. For each of age, sex and race, it also records whether the retrieved evidence supports a clinical effect on the disease. When a claim has no supporting passage, the agent writes its own query.
  3. The system issues a fixed query about differential diagnoses. The Differential Agent then adds the findings and mechanisms of competing diagnoses, and it can also write its own queries.
  4. The system issues fixed queries about acquisition and image quality. The Nuisance Modeling Agent then adds nuisance factors, such as the projection, and links each of them to the findings whose scores it can distort. For each of age, sex and race, it records whether the evidence supports an effect on how the findings appear in the image.
  5. The Graph Audit Agent compares each node and edge with the passage that it cites. It removes or corrects content with an invalid citation or a contradicting passage, but it cannot add medical content or change roles. A claim that the excerpt neither supports nor contradicts is recorded as unresolved and kept, because the excerpt that the agent sees is truncated.
  6. In one negotiation round, each agent can raise up to two challenges against content that another agent created. The agent that created the content accepts a challenge and proposes a repair, rebuts it with evidence already cited in the graph, or marks it unresolved; rebutted and unresolved content is kept. The accepted repairs are applied to a copy of the graph, and the copy replaces the graph only if it passes the compiler. Otherwise, the previous graph is kept.

In steps 2 to 5, an agent sends its candidate graph to the compiler, a deterministic program that checks the types of the nodes and edges, their roles, the required content and the citations. The agent corrects any errors and submits only a candidate that has passed. Every node and edge cites a retrieved passage.

After construction, the concepts are tested and refined (Section 3.2), and the system retrieves further passages when a concept is rewritten or replaced.

Table R6: Retrieval settings.

ItemProtocol
Results returned per searchThe 5 best passages for a template or agent query, and the 3 best for a fixed query (steps 3 and 4) or a refinement query.
Ranking and filteringPassages are ranked by BM25 score, with no score threshold and no reranking model. Duplicate passages are removed.
Stopping criterionA step ends when the agent submits a graph that has passed the compiler. Each agent call is limited to 10 tool uses, 4 searches and 2 retries after a compiler error. Only 1 of 33 chest X-ray calls used all 4 searches.
Coverage of findings and differential diagnosesThe compiler accepts a graph only if it has at least 3 direct findings of at least 2 kinds (for example, morphology and location), at least 2 differential findings, and at least 3 nuisance factors from at least 2 families.

As shown, all retrieval settings are fixed in advance, and the coverage of findings and differential diagnoses is checked by the compiler. This part has been incorporated into Appendix B.2 of our revised paper, and [1] and [2] are now cited in Section 3.1.

[1] Y. Yang et al. A Textbook Remedy for Domain Shifts: Knowledge Priors for Medical Image Analysis. NeurIPS. 2024.

[2] G. Xiong et al. Benchmarking Retrieval-Augmented Generation for Medicine. Findings of ACL. 2024.

AuthorsOpenReview ↗

Response to Question 1(ii)

Question 1(ii): Search records that link the retrieved evidence to the final graph.

Thank you for the suggestion. As requested, we summarize the search records of the released graphs in Table R7, and we follow one record from the query to the graph.

Table R7: Evidence in the released graphs.

ModalityPassages retrievedQueriesPassages from template / fixed / agent / refinement queriesPassages cited in the graphsPMC articles cited
Chest X-ray756183278 / 66 / 160 / 252222149
Skin3047784 / 18 / 55 / 1478766

Search records are kept for every claim in the released graphs. Each node and edge lists the passages that it cites, and each passage is stored with its query, its rank and score, and its source article and section. For example, the Graph Initialization Agent wrote the query "chest radiograph atelectasis radiographic signs volume loss fissure displacement elevated diaphragm linear opacity". The passage returned at rank 3 is from the section "Atelectasis" of PMC2714572 and reads "... displacement of interlobar fissures and elevation of the hemidiaphragm towards areas of atelectasis". The Atelectasis graph cites this passage for the relationship that the finding "hemidiaphragm elevation" supports the diagnosis.

This part has been incorporated into Appendix B.2 of our revised paper, where Figure 4 shows this record together with the other passages returned by the same query. The full search records will be released together with the source code.

AuthorsOpenReview ↗

Response to Weakness 1(ii) and Question 2(i)

Weakness 1(ii) & Question 2(i): How nodes and edges with unresolved disagreements are used, and why they are retained.

Thank you for the constructive comment. A challenge is left unresolved when the cited evidence can neither confirm nor refute it, because the agents are shown only the first 700 characters of each retrieved passage during negotiation. In that case, the challenged node or edge is kept, because it was added with a cited passage from the medical literature, and removing it without evidence against it could remove clinically meaningful content from the graph. The kept content is then used in all later steps like any other content: nodes with reader questions are annotated and their readers are trained and audited, and nodes and edges enter diagnostic inference with their assigned roles. Empirically, in our experiments, no challenge eventually remaines unresolved. This has been incorporated into Appendix B.2 of our revised paper.

Reviewer_yP5DOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

GC-NCE constructs literature-grounded clinical graphs and uses typed roles to constrain diagnosis from predicted concepts: direct findings and target mechanisms provide support, differential evidence provides inhibition, and anatomy gates visibility. Nuisance correction is restricted to graph-declared edges and enabled only when source-validation worst-group accuracy improves. On ten correlation-reversal benchmarks, the reported mean OOD AUROC is 78.5% for chest X-rays and 84.5% for skin lesions.

Critical strengths
  1. The clinical roles change the permitted computation rather than merely supplying concept names. Equations (6)-(8) restrict support and inhibition, and the matched-head experiments provide evidence for this design. With correction disabled and the same graph-bound readers, graph inference reaches 76.7% chest X-ray AUROC versus 66.7% for a linear head under the same training budget and selector. The corresponding skin-lesion difference is smaller, 83.8% versus 83.2% (Tables 2 and 11). The graph-content controls go beyond removing components. Role swaps and reader permutations retain the node set, parameter counts, inference components, and training/selection protocol. Each control uses three perturbation draws and three training seeds. Their substantial performance drops support the importance of the original role assignments and reader bindings (Table 12).

  2. The correction analysis acknowledges a real limitation: nuisance-predictable variation can include clinical signal (Appendix B.5). Tables 3 and 13 show both benefits and failures, including a 7.0-point AUROC loss from always correcting ISIC-site. Adaptive selection rejects correction on all three seeds for ISIC-site, ISIC-hospital, and NIH-age. This is useful evidence for treating correction as a task-dependent decision.

Critical weaknesses
  1. The headline comparison combines differences in the diagnostic method and model selection. Appendix E explicitly uses overall-validation checkpointing for primary baselines, whereas GC-NCE uses worst-group checkpointing. The matched-head controls above establish a graph-inference benefit, particularly on X-rays, but do not quantify how much of the Table 1 advantage over KnoBo, LaBo, or the other primary baselines survives a common selector. Appendix J already provides relevant evidence: its final-bank X-ray AUROCs are 72.6% without correction and 75.3% with correction under overall-validation selection, compared with 76.7% and 78.3% in the worst-group-selected correction study. The relationship between these runs deserves explicit explanation before attributing these differences entirely to selection.

The worst-group selector is specified precisely, but its effective sample sizes are unclear. It chooses the nuisance reader most associated with source diagnosis and crosses its median stratum with diagnosis; the median split need not balance the groups within each diagnosis. Validation sets contain 260-1,000 images (Table 7), yet actual proxy-group counts and correction-arm margins are not reported. Table 13 already shows mostly consistent arm choices across training seeds, with ISIC-sex selecting correction on two of three seeds. Validation-resampling results would clarify the stability of the correction-arm and checkpoint choices.

  1. Preparation uncertainty and clinical validity remain separate limitations. Appendix K evaluates five LLMs and reports diagnostic-training seed SD no larger than 0.2 points, but each backbone has only one preparation trajectory per modality. X-ray AUROC ranges from 61.9% to 78.5%; changing the LLM changes graphs, weak labels, readers, and proxy groups together, so this does not measure repeat-preparation variability. The citation checks, report-grounding rules, and reader audits are useful safeguards, but agreement with LLM-derived report labels does not independently establish image-level concept accuracy or clinical correctness of graph roles.

I lean toward acceptance on the strength of the matched-reader and graph-content controls.

Questions for authors
  1. Can the primary-baseline checkpoints be compared using the same proxy groups and worst-group selector? Please also explain whether the final-bank runs in Tables 17-18 differ from Tables 2-3 only in selection.

  2. Which nuisance reader defines the groups for each dataset, what are the label-by-stratum counts, and how large are the selected arms' worst-group accuracy margins? Is validation-resampling stability available beyond Table 13's training-seed choices?

  3. Is any independent repeat-preparation or clinician audit available? Please distinguish training-seed/backbone variation from repeat-preparation variability, and automated grounding audits from independent clinical validation.

4 comments
AuthorsOpenReview ↗

Response to Weakness 1(i) and Question 1

Weakness 1(i) & Question 1(i): The gain may come from different checkpoint-selection rules.

Thank you for raising this point. As suggested, we re-selected the checkpoints of every baseline in Table 1 with the same proxy groups and worst-group selector as GC-NCE, and we also re-selected GC-NCE with the baselines’ overall-accuracy rule (Table R1). Results show that:

  • When all methods use the worst-group selector, GC-NCE leads the best baseline by 24.0 points on chest X-rays and 5.2 points on skin lesions, and is best on all ten datasets.
  • When all methods use overall accuracy, the lead is 15.1 points on chest X-rays and 2.2 points on skin lesions, where GC-NCE is best on 8 of 10 datasets.

This part has been incorporated into Section 4.2 and Appendix E.1 of our revised paper.

Table R1: Mean OOD AUROC (%) when all methods use the same selectors.

MethodX-ray, overallX-ray, worst-groupSkin, overallSkin, worst-group
KnoBo57.254.575.575.5
LaBo44.344.879.279.3
LSL48.650.859.158.7
CIBM (IB-B)29.428.977.577.0
CIBM (IB-E)28.026.477.276.2
Linear probe27.326.450.749.7
PCBM-h26.625.050.450.0
MCBM26.024.847.245.9
C²BM24.524.344.443.0
DenseNet-12115.915.750.550.2
ViT-L/1415.515.242.442.7
GC-NCE72.478.581.584.5
Gap to best baseline+15.1+24.0+2.2+5.2

Question 1(ii): Do Tables 17–18 differ from Tables 2–3 only in selection?

Thank you for the question. Yes, they differ only in the selection rule. They use the same final bank, readers, seeds and training. Tables 17–18 select by overall validation accuracy, and Tables 2–3 select by worst-group accuracy. When we retrain the models of Tables 2–3 with the overall-accuracy rule, we obtain exactly the final-bank rows of Tables 17–18. This clarification has been added to Appendix J of our revised paper, where Tables 17–18 are numbered Tables 27–28.

AuthorsOpenReview ↗

Response to Weakness “worst-group selector” and Question 2(i)

Weakness 2(i) & Question 2(i): Group sizes and margins of the worst-group selector.

Thank you for the constructive comment. As suggested, we report the grouping reader, the group sizes and the margins for every dataset in Table R2. Groups are defined as in Appendix B.7. The images are split at this reader's median score and crossed with the two classes, which gives four groups. The margin W1 − W0 is the gap between the worst-group accuracy with correction and that without correction, and a positive margin selects correction. This part has been incorporated into Appendix B.7 of our revised paper.

Table R2: Grouping reader, number of validation images per group, and margin W1 − W0 in accuracy points (seeds 42 / 7 / 123). Class 0 and class 1 follow the order in Table 7 (Table 16 in our revised paper).

DatasetGrouping reader (AUROC)Number of images in each groupMargin W1 − W0
NIH-sexdevice or artifact over the lungs (0.667)135 / 115 / 100 / 150+2.6 / +2.6 / +1.7
NIH-agebody habitus, soft tissue (0.185)71 / 179 / 190 / 60−2.8 (all)
NIH-poscontrast or scatter haze (0.986)138 / 12 / 8 / 142+1.4 / +1.2 / +1.4
CheXpert-raceportable bedside AP (0.871)197 / 53 / 53 / 197+7.5 (all)
NIH-CheXpertcollimation or cropping (0.752)345 / 155 / 165 / 335+1.9 / +0.6 / +0.6
ISIC-sexlow lesion-skin contrast (0.847)120 / 30 / 22 / 1280.0 / +2.7 / +2.7
ISIC-agelow lesion-skin contrast (0.935)128 / 22 / 20 / 130+11.4 (all)
ISIC-sitedarker surrounding skin (0.833)214 / 86 / 69 / 231−1.2 (all)
ISIC-colorlow lesion-skin contrast (0.925)113 / 17 / 17 / 113+5.9 / +11.8 / +11.8
ISIC-hospitaldarker surrounding skin (0.944)430 / 70 / 64 / 436−1.7 (all)
AuthorsOpenReview ↗

Response to Weakness “validation resampling” and Question 2(ii)

Weakness 2(ii) & Question 2(ii): Is the selection stable under validation resampling?

Thank you for the suggestion. As suggested, we resampled the validation set 1,000 times with replacement within each group, and we repeated the checkpoint and arm selection each time. The results are summarized in Table R3. As shown, GC-NCE is stable for most datasets under validation resampling: its mean OOD AUROC is 76.5 ± 1.8 on chest X-ray and 83.4 ± 0.9 on skin. Even at the 5th percentile (72.7 and 81.9), it outperforms the strongest baseline in Table R1 (57.2 and 79.2), so GC-NCE remains the best method in both modalities. This part has been incorporated into Section 4.2 and Appendix B.7 of our revised paper.

Table R3: OOD AUROC (%) under validation resampling (mean ± SD over 1,000 resamples, with the 5th and 95th percentiles). Correction enabled is the percentage of resamples in which the selector turns on correction, averaged over the three seeds.

DatasetCorrection enabled (%)Resampling: mean ± SD [5th, 95th]
NIH-sex7986.5 ± 0.3 [86.3, 86.9]
NIH-age2753.5 ± 1.3 [52.0, 55.6]
NIH-pos2868.8 ± 8.8 [52.1, 77.2]
CheXpert-race9989.5 ± 0.1 [89.4, 89.7]
NIH-CheXpert5784.0 ± 1.8 [81.5, 86.1]
ISIC-sex4588.3 ± 0.2 [87.9, 88.5]
ISIC-age9880.2 ± 0.4 [79.5, 80.5]
ISIC-site4174.6 ± 3.9 [68.7, 78.1]
ISIC-color7686.4 ± 0.8 [84.8, 87.0]
ISIC-hospital4487.4 ± 1.5 [85.0, 89.0]
X-ray mean–76.5 ± 1.8 [72.7, 78.5]
Skin mean–83.4 ± 0.9 [81.9, 84.5]
AuthorsOpenReview ↗

Response to Weakness 3 and Question 3

Weakness 3 & Question 3: Independent repeat-preparation variance and clinical validity.

Thank you for the constructive comment. As suggested, we repeated the full preparation two more times for each of the four alternative backbones of Appendix K (GPT-5.5 is excluded due to budget constraints), and we evaluated every preparation with the main protocol (10 datasets × 3 seeds). The results are summarized in Table R4, where the value after ± is the SD over the three training seeds and the last column gives the mean and SD across the three preparations.

Table R4: Mean OOD AUROC (%) with adaptive correction for three independent preparations of the same LLM. Preparation 1 is the Appendix K run.

LLMModalityPreparation 1Preparation 2Preparation 3Across preparations
DeepSeek V4.1 FlashChest X-ray67.3 ± 0.267.3 ± 0.065.7 ± 0.566.7 ± 0.9
DeepSeek V4 ProChest X-ray67.0 ± 0.173.2 ± 0.271.3 ± 0.270.5 ± 3.2
GLM-5.3 FlashChest X-ray61.9 ± 0.168.6 ± 0.070.1 ± 0.166.9 ± 4.3
Qwen3.8 FlashChest X-ray75.6 ± 0.172.4 ± 0.270.9 ± 1.873.0 ± 2.4
Strongest baseline (Table 1)Chest X-ray–––57.2
DeepSeek V4.1 FlashSkin82.5 ± 0.280.3 ± 1.278.2 ± 0.180.3 ± 2.1
DeepSeek V4 ProSkin81.0 ± 0.280.9 ± 0.183.3 ± 0.581.8 ± 1.4
GLM-5.3 FlashSkin82.4 ± 0.183.3 ± 0.084.0 ± 0.183.2 ± 0.8
Qwen3.8 FlashSkin78.0 ± 0.182.3 ± 0.083.6 ± 0.081.3 ± 2.9
Strongest baseline (Table 1)Skin–––79.2

Preparation variability is larger than training-seed variability: the SD across preparations is 0.8–4.3 points, against 0.0–1.8 points across seeds. On chest X-rays, all twelve preparations exceed the strongest baseline (lowest 61.9 vs. 57.2), suggesting that the gain is robust to repeat preparation. This part has been incorporated into Section 4.4 and Appendix K.1 of our revised paper.

Regarding clinical validity, we tested the chest X-ray readers, which are trained on MIMIC-CXR with our report labels, against two kinds of labels that do not come from an LLM: the radiologist labels of the 202 frontal images of the CheXpert validation set [1], and the view position (AP or PA) recorded in the DICOM header of the NIH and CheXpert images. The results are summarized in Table R5. As shown, the 7 readers for the five findings that the radiologists labeled reach a mean AUROC of 0.86. The 12 readers of the 11 graphs that ask whether an image is a portable AP view separate AP from PA images with a mean AUROC of 0.94 on NIH and 0.90 on CheXpert. This part has been incorporated into Appendix B.3 of our revised paper.

Table R5: Mean AUROC of the concept readers evaluated against labels independent of LLM annotations.

Label sourceImagesReadersMean AUROC
Radiologists (CheXpert valid)202effusion, cardiomegaly, consolidation, lung opacity, device (7)0.86
DICOM view (AP vs. PA)10,723 NIHportable AP view (12)0.94
DICOM view (AP vs. PA)8,456 CheXpertportable AP view (12)0.90

[1] J. Irvin et al. CheXpert: A large chest radiograph dataset with uncertainty labels and expert comparison. AAAI. 2019.

Comments on the paper

3 comments
Senior_Area_ChairsOpenReview ↗

Open Review

This paper proposes to improve the ability of LLMs to summarize medical imaging using medical terminology, for contribution to diagnosis. It does a nice job motivating the problem by distribution shifts across hospitals/care systems, identifying the challenges, proposing solutions, and evaluating based on accuracy in predicting diagnoses.

Strengths:

  • Excellent motivation and clear, compelling introduction

  • Identifies challenges of distinguishing diagnosis-driving medical terms from less important ones for a given case, and nuisance leakage (I like that term) of non-diagnosis-driving and even irrelevant terms

  • Addresses the very important problem of producing human-comprehensible output for medical applications

  • To my knowledge the use of a clinical evidence graph to distinguish between roles of different evidence (medical terms) in a case is innovative

  • Reasonable architecture for coordinating four LLM-based agents to construct this graph

  • Extensive experiments with sound methodology including ablation studies to evaluate the value of different components

Weaknesses:

  • I'd like to see a more extensive review of, and perhaps comparison against, related approaches in summarizing medical images for clinical diagnosis. There is review of concept bottleneck work and nuisance correction work that serves as more of a background than list of potential competing systems. What is the closest competition system for this task?

Thank you for the positive assessment and the constructive comment. In short, the closest competing system is KnoBo, and we compare GC-NCE with it on all ten datasets. We have also added a more comprehensive review of related approaches.

1. The closest competing system.

The closest competing system is KnoBo (NeurIPS-24) [1]. It studies the same problem of interpretable, concept-based diagnosis under domain shift, and we use its ten confounded datasets and its frozen image encoders. Both methods select concepts from the medical literature but combine them differently. KnoBo combines a flat list of concepts with a linear head, whose weights are learned from the source data under a prior from the literature. GC-NCE assigns each concept a clinical role in an evidence graph, such as a finding or mechanism that supports the diagnosis, a finding of a competing diagnosis, anatomy, or a nuisance factor. It computes the diagnosis by probabilistic inference that follows these roles. It corrects finding scores only along the nuisance influence edges, and only when source validation supports the correction.

The next closest systems are LaBo (CVPR-23) [2], the strongest baseline on skin lesions, and C²BM (NeurIPS-25) [3], which predicts through a concept graph learned from the source data. Table R15 compares the four systems.

Table R15: The closest competing systems under the protocol of Table 1. OOD AUROC (%) is the mean over the five datasets of each modality, and Δ = ID − OOD.

SystemConceptsHow the diagnosis uses the conceptsChest X-ray OOD / ΔSkin OOD / Δ
KnoBoSelected from the medical literatureLinear head with priors from the literature57.2 / 38.175.5 / 17.5
LaBoPubMed concepts released with KnoBoLinear head44.3 / 51.379.2 / 10.3
C²BMSame concept readers as GC-NCEGraph learned from the source data24.5 / 73.644.4 / 53.0
GC-NCEClinical evidence graph built from the literatureRole-constrained inference with nuisance correction78.5 / 10.684.5 / 5.1

As shown, GC-NCE has the highest OOD AUROC and the smallest domain gap in both modalities.

2. A more comprehensive review of related approaches.

We group the related approaches into seven families in Table R16.

Table R16: Families of related approaches.

FamilyRepresentative workOutputHow the diagnosis is formed
Radiology report generationR2Gen (EMNLP-20), R2GenCMN (ACL-21), PromptMRG (AAAI-24), MAIRA-2 (arXiv-24)Free-text reportStated in the generated report
Knowledge-graph-guided report generationZhang et al. (AAAI-20), DCL (CVPR-23)Free-text report guided by a graph of findingsStated in the generated report
Medical vision–language modelsLLaVA-Med (NeurIPS-23), Med-PaLM M (NEJM AI-24), BiomedGPT (Nat. Med.-24)Free-text answer or reportStated in the generated text
Retrieval-augmented and agentic systemsMMed-RAG (ICLR-25), MedAgent-Pro (ICLR-26)Free-text reasoning with retrieved knowledgeDecided by a language model at test time
Group-robust trainingGroupDRO (ICLR-20), JTT (ICML-21), DFR (ICLR-23)Class predictionGroup-robust objective, reweighting, or last-layer retraining
Concept bottleneck modelsCBM (ICML-20), LaBo (CVPR-23), KnoBo (NeurIPS-24)Scores of named conceptsLearned weights over the concept scores
Concept graph modelsCausal CGM (ICLR-25), C²BM (NeurIPS-25)Concept scores and a concept graphThrough a graph learned from the source data
GC-NCE (ours)–Concept scores with clinical rolesThrough a clinical evidence graph built from the literature before training

As shown, the first four families describe each test image in free text, and the explanation is text generated together with the diagnosis. Group-robust training addresses the same kind of shift by changing how the classifier is fitted, but it does not show the evidence behind a prediction. GC-NCE belongs to the concept-based families: it computes the diagnosis from the named clinical concepts that it reports, so these concepts, with their roles and scores, are the evidence that the model uses. It differs from ordinary concept bottleneck models in that its graph fixes the role that each concept can play in the diagnosis.

This part has been incorporated into Section 2 and a new Appendix L (Tables 35–36) of our revised paper, where the full references for Table R16 are given.

[1] Y. Yang et al. A Textbook Remedy for Domain Shifts: Knowledge Priors for Medical Image Analysis. NeurIPS. 2024.

[2] Y. Yang et al. Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification. CVPR. 2023.

[3] G. De Felice et al. Causally Reliable Concept Bottleneck Models. NeurIPS. 2025.

AuthorsOpenReview ↗

Summary of Rebuttal

We sincerely thank the Senior Area Chair and the reviewers for their time and constructive feedback. We are glad that the core idea of the paper was well received. The Senior Area Chair finds the use of a clinical evidence graph to distinguish the roles of different evidence "innovative". Reviewer yP5D notes that the clinical roles "change the permitted computation rather than merely supplying concept names", and Reviewer aqKZ sees this design as "a reasonable alternative to classifiers that arbitrarily exploit correlations among conceptual scores".

We have addressed all comments in our individual responses. Below, we summarize the main changes, all of which are marked in blue in the revised paper.

  • Same checkpoint selector (yP5D): We re-selected every baseline with GC-NCE's worst-group selector, and we re-selected GC-NCE with the baselines' overall-accuracy rule. GC-NCE has the highest mean OOD AUROC in both modalities under either rule (Section 4.2, Appendix E.1).
  • Selector groups and validation resampling (yP5D): We reported the grouping reader, the group sizes and the margins of every dataset. When the validation set is resampled 1,000 times, GC-NCE remains the best method in both modalities (Section 4.2, Appendix B.7).
  • Repeat preparations and LLM settings (yP5D, aqKZ): We ran three independent preparations for each of the four alternative LLMs. The variability across preparations is larger than across training seeds, but for every LLM the mean over its three preparations is still above the strongest baseline in both modalities. We also measured how much the graphs change between preparations, and we listed the top-p setting and reasoning effort of each LLM (Section 4.4, Appendix K.1).
  • Concept labels and readers (yP5D, aqKZ): We compared our report labels with two rule-based labelers, and we measured the effect of added false negatives and the agreement between two annotations by the same LLM. We also tested the readers against radiologist labels and the DICOM view position (Appendix B.3).
  • Graph construction and retrieval (aqKZ): We described the six construction steps, the retrieval settings and the search records. We also explained why content with an unresolved challenge is kept and how it is used (Section 3.1, Appendix B.2).
  • Reader gate (aqKZ): We repeated the refinement under five other thresholds from 0.55 to 0.80. The OOD AUROC does not increase with the threshold, while a higher threshold requires more annotations (Appendices B.3 and J.1).
  • Related approaches (Senior Area Chair): We identified KnoBo (NeurIPS-24) as the closest competing system and compared GC-NCE with KnoBo, LaBo (CVPR-23) and C²BM (NeurIPS-25). We also added a review of seven families of related approaches (Section 2, Appendix L).

Full results are provided in Tables R1–R16 of our individual responses. Once again, we thank the Senior Area Chair and the reviewers for their constructive comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.