acceptodds
Under review as a conference paper at ICLR 2027

CLAIR-Med: Selective Neural–Symbolic Reasoning with Weakly Supervised Concepts for Medical Image Classification

Abstract

Rule-based concept models expose an explicit decision process but are often less accurate than neural predictors. However, always falling back to the neural model renders the symbolic explanation irrelevant to the final decision. We investigate a narrower question: how often can an explicit symbolic branch make the prediction while retaining competitive task utility? CLAIR-Med constructs medical concept vocabularies using a language model, derives pseudo concept targets from vision–language similarity, and learns an image-to-concept interface without manual concept annotations. The predicted concept state is consumed by two structurally different branches: a concept-augmented predictive branch and a Deep Concept Reasoner (DCR) branch that executes fuzzy rules on concept truth values. We distill class-level predictive information from the predictive branch into the DCR branch and use the predictive entropy of DCR's output distribution to dynamically route each input between the two branches at inference. Experiments on five medical image datasets and CUB show that distillation narrows the predictive–DCR utility gap across all six benchmarks without degrading predictive-branch accuracy. Entropy-based routing enables the DCR branch to supply the final prediction on a subset of inputs while retaining near-predictive task utility. We further evaluate whether the displayed rules reproduce DCR's decision behavior, while blinded medical-domain ratings and concept-deletion analysis assess concept quality and prediction sensitivity, respectively. Together, these results demonstrate the feasibility of selective symbolic prediction through an explicit concept-rule pathway while retaining competitive task utility.

Primary area:
interpretability and explainable AI
Keywords:
Medical Image Classification, Interpretable Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_W7eZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CLAIR-Med proposes a neural symbol medical image classification framework that eliminates the need for manual concept annotation, aiming to alleviate the trade-off between the interpretability of explicit symbolic reasoning and the predictive performance of neural predictors. The paper conducts experiments on five medical image datasets and the CUB dataset, experimental results show that the framework achieves high symbol prediction coverage while maintaining near-neural predictor accuracy. However, existing results still lack sufficient validation regarding the independent contribution of the key routing mechanism and the clinical reliability of the learned concepts.

Critical strengths

1.The research question has clear relevance and practical significance. This paper focuses on whether explicit symbolic reasoning can truly participate in final decision-making, shifting the research focus from the traditional trade-off between prediction accuracy and interpretability to selective symbolic prediction with clear decision-making significance, particularly applicable to applications such as medical imaging that emphasize decision interpretability. 2.Experiments cover multiple dimensions and provide preliminary evidence supporting the framework's effectiveness. The authors validated the method on six datasets and conducted analyses from the perspectives of symbolic branching accuracy, knowledge distillation, dynamic routing, rule replay consistency, and concept intervention. Furthermore, concept deletion experiments and medical expert evaluations provide some evidence for the potential effectiveness of concept representation. These results demonstrate that the proposed framework has practical potential, rather than merely remaining at the conceptual system design level.

Critical weaknesses

1.The independent contributions of the core technological innovations are not yet clear, and existing experiments are insufficient to fully demonstrate the superiority of the uncertainty-driven routing mechanism. This paper integrates language model-assisted concept generation, visual-language pseudo-supervision, concept bottleneck learning, deep concept reasoning, knowledge distillation, and uncertainty-driven selective prediction into a unified framework. However, these component technologies are all closely related to existing research, and the paper needs to more clearly explain its substantial innovation compared to existing concept models and neural symbol selective prediction methods. Furthermore, selective symbol prediction is one of the core contributions of this paper, but existing experiments mainly demonstrate that the routing model can achieve high symbol coverage while maintaining overall accuracy, without fully demonstrating that the proposed uncertainty-driven routing strategy has a significant advantage over standard confidence routing or other selective classification strategies. 2. The clinical reliability of medical concepts and the standardization of imaging experimental demonstrations still have shortcomings. This paper uses language models to generate medical concepts and constructs pseudo-concept supervision through visual-language similarity, thereby avoiding expensive manual concept annotation. However, this process also introduces potential semantic bias and visual localization errors. Concepts generated by language models may not necessarily correspond to identifiable clinical signs in specific medical images, and high similarity obtained by visual language models cannot be directly equated with the actual existence of clinical concepts. Although the authors provided expert evaluations and concept deletion experiments, these experiments primarily verified the subjective rationality of concepts and their impact on prediction probabilities, and are insufficient to prove the clinical accuracy of concept predictions. Furthermore, the creation of some figures and the presentation of medical images in the paper are not rigorous enough. Particularly in the presented CT images, different examples seem to lack unified or clearly defined window width and window level settings. For research involving medical image analysis, image display standards are not only a typesetting issue but also relate to the reliable interpretation of qualitative results. The authors should clearly define the image display and preprocessing protocols and standardize the revisions of relevant figures and tables.

Questions for authors
  1. Independent Contribution to Uncertainty-Driven Routing Could the authors further explain the substantial advantages of the proposed DCR uncertainty routing mechanism compared to standard maximum prediction probability, prediction entropy, classification margin, or other selective classification strategies?
  2. Usability and Clinical Accuracy of Pseudo-Concepts This paper utilizes a language model to generate a medical concept vocabulary and obtains pseudo-concept labels through visual-linguistic similarity. However, the semantic plausibility of a concept does not necessarily imply its verifiable clinical authenticity in medical images. How do the authors strike a balance between reducing the cost of manual concept annotation and ensuring the clinical accuracy of concepts? Specifically, are there concept-level reference annotations verified by medical experts to quantitatively evaluate the Precision, Recall, or F1-score of concept predictions? For concepts generated by the language model but not reliably identifiable from images, are effective screening, calibration, or elimination mechanisms designed? Furthermore, how do these errors propagate to the symbolic reasoning and final prediction of DCR when concept predictions are inconsistent with clinical facts?
2 comments

Thank you for the careful review. We appreciate your constructive questions about selector quality, concept reliability, error propagation, and medical-image display.

Entropy versus other selectors. The coverage-matched table fixes the checkpoint, branch predictions, and DCR coverage within each seed, so it compares selector ranking rather than changing the fraction of symbolic assignments. Our contribution is to use DCR entropy to decide when the explicit rule branch supplies the final prediction. Margin performs better on BloodMNIST at 80% coverage and Chest X-Ray at 50%, and random routing also exceeds entropy at some high-coverage points. Among the three confidence-based selectors (MSP, margin, and entropy), entropy ranks first or ties for first at 10 of the 12 reported operating points. At 80% coverage it exceeds margin by 0.41 points on Chest X-Ray and 0.18 points on Fitzpatrick17k. For binary COVID-CT, entropy, MSP, and margin induce the same ranking up to ties. We therefore do not claim uniform or statistically significant superiority. The retrospective oracle remains higher, leaving headroom for better branch selection.

Concept correctness and filtering. Our annotation-free claim concerns training supervision; it does not mean that concept correctness is untestable. On the SkinCon-annotated subset of the Fitzpatrick17k test split, we evaluated concept-head predictions against expert per-image annotations. Across the 19 mapped concepts, macro precision, recall, and F1 were 0.73, 0.75, and 0.74, respectively (Table B). The visual-screening ablation covers four datasets and uses Qwen3-VL-32B-Instruct, a class-stratified 10% subset of training images, no validation/test images, removal only when a concept is judged absent in every sampled image, and five-seed retraining. Concepts removed and original/after-screening routed accuracy are: BloodMNIST 3 and 97.87/97.88; OCT2017 5 and 99.79/99.72; COVID-CT 1 and 92.06/92.21; CUB 6 and 79.41/79.40. The changes range from -0.07 to +0.15 percentage points. This four-dataset ablation is not claimed for Chest X-Ray or Fitzpatrick17k. The filter removed only concepts that the VLM judged absent from every sampled training image.

Table A — Macro AUROC

DatasetHead vs CLIP pseudo-labelsCLIP similarity vs official annotationsConcept head vs official annotations
CUB0.850.800.77
SkinCon-annotated Fitzpatrick17k subset0.870.830.82

Table B — Concept-head metrics against official annotations

DatasetMapped model conceptsMacro precisionMacro recallMacro F1
CUB860.720.780.75
SkinCon-annotated Fitzpatrick17k subset190.730.750.74

Table S — Visual-screening ablation

DatasetConcepts removedOriginal routed accuracy (%)After screening accuracy (%)
BloodMNIST397.8797.88
OCT2017599.7999.72
COVID-CT192.0692.21
CUB679.4179.40

Error propagation. Concept errors can propagate through DCR literals to class scores and routing. Relevance filtering and entropy fallback provide partial mitigation only: a confidently wrong DCR output can pass the gate, and the predictive branch also consumes concept information. Existing deletion and intervention analyses measure sensitivity and functional editability, not clinical correctness. Appendix H.3 now states this limitation explicitly.

Medical-image display. We standardized panel dimensions, annotation boxes, labels, and bar scales. We preserved image aspect ratios and applied no additional brightness or contrast adjustment. The COVID-CT dataset README states that the original Hounsfield unit (HU) values were lost when CT images were reproduced in publications. We therefore could not apply a common HU-based window width and level to the released images shown in our figures. The examples should therefore be read as retrospective methodological illustrations, not evidence of diagnostic equivalence or clinical readiness.

In the newly submitted version of the manuscript, the revisions highlighted in blue include the image-level concept evaluation in Appendix T.4 (Tables 27 and 28), the visual-screening ablation and discussion of error propagation in Appendix H.3 (Table 9), and the clarifications on COVID-CT preprocessing and image display in Appendices G.1 and Q, respectively. The existing coverage-matched selector comparison is in Appendix V (Table 31).

Reviewer_BVbZOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CLAIR-Med, a selective neural–symbolic framework for medical image classification that lets an explicit concept-rule model make the prediction whenever it is confident and falls back to a neural model otherwise.

GPT-4o generates a concept vocabulary, and thresholded CLIP image–text similarity provides concept pseudo-labels that a concept head learns to predict. The predicted concept scores feed two paths: a neural classifier that also uses the image features, and a Deep Concept Reasoner (DCR) that applies a learned fuzzy rule per class. The neural path is distilled into DCR, and at inference DCR’s prediction is used when its entropy is below a validation-tuned threshold.

On five medical datasets and CUB-200, the method reports the highest accuracy among the compared baselines. Distillation narrows the accuracy gap between the two paths, and 78–99% of test inputs are routed to DCR. A blinded study with three medical-domain evaluators rates its concepts above LF-CBM’s.

Critical strengths
  1. The distillation part of the second contribution is well supported. Distilling the predictive branch into DCR improves DCR accuracy on all six datasets without hurting the predictive branch. On the harder datasets the gap between the two branches roughly halves, e.g., 4.30 → 1.89 on Fitzpatrick17k (Table 20, paired over five seeds).

This result shows that the distillation component works as claimed.

  1. The evaluation is broad. The method is tested on five medical datasets of varying difficulty (2 to 114 classes) plus CUB, with accuracy, F1, and AUROC, five seeds, and backbone and pseudo-labeler variations (Tables 1, 12–18, 24).

This breadth is an important component of a rigorous evaluation.

Critical weaknesses
  1. The novelty is unclear. The authors should state precisely what is new in this paper, especially relative to two prior works that are not cited:

Hybrid Rule Sets (Wang, ICML 2019) already defines the problem. An interpretable model predicts the inputs it covers, a black box handles the rest, and “transparency” (the fraction routed to the interpretable model) is the same quantity as symbolic coverage. Route, Interpret, Repeat (Ghosh et al., ICML 2023) already proposes the approach for concept-based image models, including medical images. It distills a black box into concept-based logic experts and routes inputs between them and a residual network with a learned selector.

If the problem and the approach both come from prior work, the paper’s contribution reduces to a specific combination of known components, and its own results show that the one new choice, entropy-based routing, adds little (Table 27).

  1. The concepts are not shown to be semantically grounded. The paper’s premise is that decisions can be inspected through named concepts, but it never shows that the concept scores reflect what their names say. Three checks are missing:

Agreement with CLIP. The concept head is trained to match CLIP pseudo-labels, but its agreement with them on test images is never reported. The same scores are also shaped by the classification and distillation losses, so they may encode the class rather than the concept. Correctness of CLIP’s labels. The CLIP labels themselves are never checked against experts or ground truth on medical images, even though ground truth exists for CUB and Fitzpatrick17k (SkinCon). By contrast, Route, Interpret, Repeat uses annotated concepts and keeps only those with validation AUROC ≥ 0.7. Evaluation. The rating and deletion studies measure how plausible the concepts look and how much the prediction depends on them, not whether each concept is present in the image.

Without grounding, the symbolic path is a classifier with concept names attached, and its high coverage does not make decisions inspectable, which is the paper’s main motivation.

Questions for authors
  1. Novelty. What is new relative to Hybrid Rule Sets (Wang, ICML 2019) and Route, Interpret, Repeat (Ghosh et al., ICML 2023), both in the problem definition and in the method?

  2. Concept grounding. How well do the learned concept scores agree with the CLIP pseudo-labels on test images? How accurate are the concept scores against ground truth on CUB (attribute labels) and Fitzpatrick17k (SkinCon)?

4 comments

Thank you for the careful review. We appreciate your positive assessment of the distillation results and your questions about novelty and concept grounding.

Contribution relative to prior work. We agree that the novelty delta should be concrete. VLG-CBM already studies automated visual grounding and leakage-aware evaluation; CONDA studies unlabeled target-domain adaptation; CB-LLM studies automated concept construction; and Route, Interpret, Repeat and Hybrid Rule Sets study selective or iterative interpretable experts. CLAIR-Med’s setting is the shared weakly supervised concept interface between a concept-augmented predictive branch and a single explicit DCR branch, one-way class-distribution distillation to DCR, and entropy-based branch selection. We claim this concrete combination and its coverage--utility--replay evaluation; we do not claim dual branches, distillation, medical application, or entropy as individually new. The paired multi-seed analysis supports the intended mechanism: the utility gap narrows mainly through improved DCR accuracy while predictive accuracy remains stable.

Semantic grounding. We agree that head--CLIP agreement cannot prove semantic correctness: a head can retain pseudo-supervision even when the pseudo-labeler is wrong. We therefore separate pseudo-label consistency from evaluation against independent image-level annotations on CUB and the SkinCon-annotated Fitzpatrick17k subset. Human labels are used only for evaluation, semantic mappings are fixed before metric computation, and unmatched concepts are not treated as negatives. The 98 CUB concepts were selected from the official attribute vocabulary. At a text-similarity threshold of 0.75, 86 of these concepts and 19/82 Fitzpatrick17k concepts met the automatic matching criterion for evaluation against human annotations. The three macro-AUROC columns (head vs CLIP pseudo-labels / CLIP similarity vs official annotations / concept head vs official annotations) are 0.85/0.80/0.77 for CUB and 0.87/0.83/0.82 for the SkinCon-annotated Fitzpatrick17k subset. Mapped-concept macro precision/recall/F1 are 0.72/0.78/0.75 and 0.73/0.75/0.74, respectively. Human labels are evaluation-only; P/R/F1 use the fixed concept-decision rule, and unmatched concepts are not treated as negatives. The conclusion is restricted to the evaluated mapped subset. The result does not certify the full generated vocabulary, turn CLIP into a ground-truth annotator, or establish clinical correctness. The existing blinded ratings and deletion analysis remain complementary evidence about perceived relevance and prediction sensitivity.

Table A — Macro AUROC

DatasetHead vs CLIP pseudo-labelsCLIP similarity vs official annotationsConcept head vs official annotations
CUB0.850.800.77
SkinCon-annotated Fitzpatrick17k subset0.870.830.82

Table B — Concept-head metrics against official annotations

DatasetMapped model conceptsMacro precisionMacro recallMacro F1
CUB860.720.780.75
SkinCon-annotated Fitzpatrick17k subset190.730.750.74

In the newly submitted revised manuscript, we added a brief discussion in Section 2 (pp. 2–3) explaining how CLAIR-Med relates to Hybrid Rule Sets and Route, Interpret, Repeat. We also added an image-level concept evaluation, summarized in Section 4.5 and reported in detail in Appendix T.4 (Tables 27 and 28). Both additions are highlighted in blue.

Thank you for the rebuttal and revisions.

The new image-level evaluation addresses an important gap in the original paper by comparing predicted concept scores with human annotations on CUB and SkinCon.

I still view the methodological contribution as incremental. The revised Related Work clarifies the differences from Hybrid Rule Sets and Route, Interpret, Repeat. The Introduction, however, still leads its contribution list with “We formalize selective symbolic prediction …,” making it harder to distinguish the established problem from this paper’s specific contribution. I would revise the Introduction to emphasize the weakly supervised design and its main empirical result: DCR makes the final prediction on a large fraction of test examples while overall accuracy remains close to that of the neural branch alone.

Given the new grounding evidence, I am raising my rating from 2 to 3.

We have revised the first contribution in the Introduction (Section 1, p. 2) to emphasize the weakly supervised design and the main empirical finding, with the revision highlighted in blue in the updated PDF. Thank you for your careful review, constructive suggestions, and positive reassessment of our work. Your feedback has been very helpful in clarifying our contribution and improving the presentation of the paper. We sincerely appreciate your time and effort throughout the discussion.

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

This paper proposes CLAIR-Med, a selective neural–symbolic reasoning framework for medical image classification. The method uses GPT and CLIP to construct weakly supervised concept representations, improves the predictive performance of the DCR symbolic branch through knowledge distillation, and dynamically selects between the neural and symbolic branches based on predictive entropy. Experiments on six datasets evaluate classification performance, symbolic coverage, concept quality, and rule fidelity.

Critical strengths
  1. The paper addresses the problem of increasing the use of explicit symbolic reasoning while maintaining predictive performance, which has practical value.
  2. The experiments are relatively comprehensive, and the results show that distillation effectively reduces the performance gap between the two prediction branches.
Critical weaknesses
  1. The necessity of explicit symbolic reasoning is not sufficiently validated. CLAIR-Med mainly combines existing techniques, including concept learning, DCR, knowledge distillation, and uncertainty-based routing. Although the experiments demonstrate the effectiveness of the overall framework, they do not establish the necessity of explicit symbolic reasoning compared with standard concept classifiers.
  2. The expert evaluation in Table 3 mainly assesses clinical relevance and explanation quality, while the concept deletion experiments in Table 4 measure prediction sensitivity. These results do not directly demonstrate that the predicted concepts are actually present in the corresponding medical images.
  3. The routing ablation in Appendix B uses different checkpoint selection strategies, so the performance differences may reflect changes in both model parameters and routing mechanisms. In addition, Table 2 shows that routed accuracy does not exceed the standalone neural branch on most datasets.
  4. Appendix G.3 defines static fusion as an equal-weight combination of the two branch outputs. However, Eq. (4) produces classification logits, while Eq. (7) produces fuzzy truth values in [0, 1]. Directly combining these outputs may lead to an unfair comparison due to their different scales.
Questions for authors
  1. How would the predictive performance and symbolic branch usage change if DCR were replaced with a standard concept classifier while retaining knowledge distillation and routing?
  2. Can the correctness of predicted concepts be evaluated using manually annotated medical concepts rather than relying only on clinical relevance?
  3. Is static fusion performed on outputs with the same scale? Can the routing ablation be repeated using fixed checkpoints?
2 comments

Thank you for the careful review. We address the role of DCR, concept correctness, routing isolation, and static fusion separately.

Role of DCR. DCR is used because it executes inspectable concept-level fuzzy literals and conjunctions. A standard classifier on concept scores would be a valid but different auxiliary baseline; it would remove the explicit rule computation. Our claim is therefore about a measurable symbolic pathway, not about DCR being necessary for best accuracy. In the existing five-seed analysis, distillation reduces the predictive--DCR gap from 4.30 to 1.89 points on Fitzpatrick17k, 3.25 to 1.43 on CUB, and 4.69 to 2.67 on Chest X-Ray, while predictive accuracy remains stable. Routed rule replay agreement is 91.05--98.52%. These results support symbolic participation and inspectability, not universal superiority.

Concept correctness. We agree that CLIP agreement, relevance ratings, and deletion sensitivity do not establish image-level concept correctness. They measure pseudo-supervision consistency, perceived relevance, and prediction sensitivity, respectively. We report an evaluation against independent image-level annotations on CUB and the SkinCon-annotated Fitzpatrick17k subset, using labels only for evaluation and excluding unmatched concepts from the negative class. The 98 CUB concepts were selected from the official attribute vocabulary. At a cosine-similarity threshold of 0.750.75, 86 of these concepts and 19/82 Fitzpatrick17k concepts met the automatic matching criterion for evaluation against human annotations. Macro AUROC is 0.85/0.80/0.77 on CUB and 0.87/0.83/0.82 on the SkinCon-annotated Fitzpatrick17k subset for head vs CLIP pseudo-labels / CLIP similarity vs official annotations / concept head vs official annotations. On mapped concepts, macro precision/recall/F1 are 0.72/0.78/0.75 for CUB and 0.73/0.75/0.74 for the SkinCon-annotated Fitzpatrick17k subset. AUROC is computed per concept from continuous scores and then macro-averaged. Head–CLIP comparisons use all 98 CUB and 82 Fitzpatrick17k concepts; comparisons with human annotations use only the 86 and 19 mapped concepts, respectively. For precision, recall, and F1, concept-head scores are binarized at 0.50.5, and each metric is computed per mapped concept and then macro-averaged. The conclusion remains limited to mapped concepts and evaluated images; we do not claim validation of the full vocabulary or clinical correctness.

Table A — Macro AUROC

DatasetHead vs CLIP pseudo-labelsCLIP similarity vs official annotationsConcept head vs official annotations
CUB0.850.800.77
SkinCon-annotated Fitzpatrick17k subset0.870.830.82

Table B — Concept-head metrics against official annotations

DatasetMapped model conceptsMacro precisionMacro recallMacro F1
CUB860.720.780.75
SkinCon-annotated Fitzpatrick17k subset190.730.750.74

Routing isolation. The original operating-point table changes checkpoint selection together with routing. With KD enabled, we select one checkpoint by predictive-branch validation accuracy with routing disabled. We then select the routing threshold on held-out validation data for that checkpoint and evaluate routing on and off on the same test examples (Table R1). Routing on improves over the predictive branch by 0.08--1.12 percentage points, with 77.17--98.62% of examples assigned to DCR. We explicitly distinguish this protocol from the coverage-matched selector analysis. Routing is intended to expose a controllable symbolic utility/coverage frontier; it need not outperform the predictive branch at every operating point. For example, the existing main table reports 79.41% routed accuracy on CUB at 78.30% DCR coverage versus 80.24% predictive-branch accuracy.

Table R1 — Fixed-checkpoint routing (KD enabled)

DatasetPredictive accuracy (%)DCR accuracy (%)Routed accuracy (%)DCR coverage (%)
OCT201799.5899.6099.7198.62
CUB79.0678.2479.3577.17
BloodMNIST97.6497.4197.7295.78
COVID-CT90.8990.8591.8686.82
Chest X-Ray77.0375.2878.0678.67
Fitzpatrick17k80.4579.1681.5782.66

Static fusion. Thank you for identifying the inconsistency. The submitted version misstated the fusion equation. We have corrected Appendix G.3: static fusion averages the predictive logits and the log-odds of the clipped DCR truth scores with equal weights, then predicts the class with the largest fused score. This is a score-space clarification, not a new performance claim.

In the newly submitted version of the manuscript, we highlight the revisions addressing your questions in blue. Section 4.5 points to the new concept evaluation in Appendix T.4 (Tables 27 and 28), and Section 4.8 points to the new fixed-checkpoint routing ablation in Appendix B (Table 6). The corrected static-fusion equation appears in Appendix G.3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.