Pith. sign in

REVIEW 3 major objections 4 minor 28 references

Learning Concept-Driven Logical Rules for Interpretable and Generalizable Medical Image Classification

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A concept-based medical image classifier that learns explicit Boolean rules from binary visual concepts, improving out-of-distribution accuracy while keeping predictions interpretable.

desk verdict Useful extension of concept bottlenecks with learnable logical rules, but the headline OOD advantage over hard-CBM is not cleanly attributed to logic rather than added nonlinear capacity. read the letter →

arxiv 2505.14049 v2 pith:7WSSFPHQ submitted 2025-05-20 cs.CV

Yibo Gao , Hangqi Zhou , Zheyao Gao , Bomin Wang , Shangqi Gao , Sihan Wang , Xiahai Zhuang This is my paper
classification cs.CV
keywords conceptinterpretabilitylogicalmedicalrulesclassificationconcept-baseddecision
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Medical image classifiers are often black boxes. Concept Bottleneck Models (CBMs) try to fix this by first predicting a list of human-understandable concepts (e.g., 'crust', 'erythema') and then making a diagnosis from those concepts. But if the concept predictions are soft probabilities or embeddings, the model can silently use extra information hidden inside those soft values, a problem called concept leakage.

The authors propose CRL, which binarizes concept predictions to zeros and ones, then passes the binary vector through logical layers that compute AND and OR combinations. This produces explicit rules such as 'crust AND erythema' as evidence for malignancy. A final linear layer weights which rules contribute to each class. Training uses a differentiable approximation of the logical operations and a straight-through estimator for the binarization.

On Fitzpatrick17k (skin lesions) and PBC (white blood cells), CRL matches the accuracy of existing concept-based methods. On the out-of-distribution DDI skin dataset, CRL's diagnostic accuracy drops only 2.49 points versus 9.74 for hard-CBM, supporting the claim that binary rules generalize better. The paper shows example rules that agree with clinical knowledge, but it does not quantify rule quality or compare logical layers against a generic nonlinear classifier on binary concepts.

Extended reading notes

Core claim

The paper's central assertion is that by making decisions based on domain-invariant logical rules, CRL both provides global interpretability and improves generalizability to unseen domains. Concretely, Section 3.2 reports that on the unseen DDI dataset, CRL reaches 73.46 diagnostic accuracy and a 2.49 point drop, while the best binary-concept baseline hard-CBM drops 9.74 points, attributing the gain to logical layers rather than model capacity.

Load-bearing premise

The load-bearing premise is that the logical layer architecture (two layers of 256 nodes with conjunction/disjunction operations) is responsible for the OOD robustness, and not simply the extra nonlinear capacity or another artifact of the binary-concept pipeline. The paper never compares CRL against a generic nonlinear classifier (e.g., an MLP) on the same binary concepts, so the attribution of the OOD gain to 'domain-invariant logical rules' is not isolated from capacity. This premise enters at Section 3.2, Table 2, where hard-CBM is the only binary-concept baseline.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes Concept Rule Learner (CRL), a concept-based classifier that binarizes concept activations and passes them through differentiable logical layers of AND/OR operations, followed by a linear layer. The authors claim that this design provides both local concept explanations and global logical rules, mitigates concept leakage, and improves out-of-distribution generalization on medical image classification. Experiments are reported on Fitzpatrick17k skin disease diagnosis, PBC white blood cell classification, and on the unseen DDI dermatology dataset. The central quantitative claim is that CRL drops only 2.49 diagnostic accuracy points on DDI while the best binary-concept baseline hard-CBM drops 9.74 points.

Significance. If the central generalizability claim holds, CRL would be a useful contribution to interpretable medical image classification: it couples global rule-based explanations with concept-level reasoning and appears to improve robustness to domain shift. The paper has several strengths: the code is released, experiments report standard deviations and multiple seeds, the DDI dataset is genuinely held out and not used for fitting, and comparisons include several recent concept-based baselines. However, the attribution of the OOD gain specifically to logical-rule structure is not yet established, because the comparison does not isolate logical structure from added nonlinear capacity. The interpretability analysis is also qualitative only.

major comments (3)
  1. [Section 3.2, Table 2] The claim that 'domain-invariant logical rules' drive the small OOD drop is underdetermined. CRL inserts two logical layers of 256 nodes each between the binary concept vector and the label predictor, while hard-CBM uses a linear classifier on the same binary concepts. The comparison therefore confounds logical-rule structure with added nonlinear capacity. I request a matched-capacity control: train a generic MLP (with a comparable number of parameters and FLOPs) on the same binary concept vector, and report its DDI drop and in-distribution accuracy. If a generic MLP reproduces the 2.49-point drop, the central attribution in the abstract and Section 3.2 must be revised.
  2. [Section 3.2, Table 2] The generalizability claim rests on a single unseen dataset, DDI. A title- and abstract-level claim about generalizability to unseen domains cannot be supported by one held-out dermatology dataset. I recommend adding at least one additional OOD evaluation, such as a different domain shift (e.g., a synthetic corruption or another clinically related dataset), or softening the claim to state that the improvement is shown on one unseen dataset.
  3. [Section 3.2, Model interpretability analysis; Section 2.4] Concept leakage mitigation is a central motivation of the paper, but leakage is never measured directly. The claim that CRL 'mitigates the issue of concept leakage' is inferred only from OOD accuracy differences. I suggest adding a leakage diagnostic, such as probing the binarized concept bottleneck for task-relevant information or performing an intervention test, or explicitly stating that leakage is not directly measured and tempering the corresponding claims.
minor comments (4)
  1. [Table 1, PBC block] The CEM row cites reference [8], but CEM is reference [7] in the bibliography; please correct the citation.
  2. [Section 3.1, Implementation details] The hyperparameters for the logical layers (number of layers, number of nodes, lambda) are fixed without a sensitivity analysis; a short ablation would help assess how robust the OOD result is to these choices.
  3. [Figure 3] The extracted rules are shown as examples only; please clarify how rules are selected or pruned after training, and state how many rules remain for each task under the chosen lambda.
  4. [Section 2.3, Eq. (1)] The projection function P(x) = 1/(1 - log x) is unusual; please clarify its domain and why it prevents gradient vanishing, and confirm that the continuous relaxation exactly recovers the Boolean operations when all inputs are binary.
Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central result depends on manually chosen hyperparameters (lambda, layer count, node count) and on unverified background assumptions about the logical activation approximation, the straight-through gradient, and the reliability of concept annotations. No new physical entities or theoretical constructs are introduced.

free parameters (5)
  • number_of_logical_layers = 2
    Architecture choice stated in Section 3.1; affects rule expressiveness and capacity, no sensitivity analysis.
  • nodes_per_logical_layer = 256
    Hand-chosen capacity for conjunction/disjunction nodes; affects the number R of extracted rules.
  • regularization_lambda = 5e-6
    Controls rule complexity in Eq. (2), Section 2.3; chosen without ablation or grid search.
  • concept_loss_balance = 1
    Balance between concept and task loss in Eq. (2); set to 1 without sensitivity analysis.
  • binarization_threshold = 0.5 (implied)
    The binarizer q maps concept activations to {0,1}; the threshold is not explicitly stated in the paper.
assumptions (4)
  • standard math Logical activation functions from Payani and Fekri (2019) accurately approximate Boolean AND and OR in continuous training
    Section 2.3 uses Conj and Disj from ref. [19]; the equivalence holds only for binary inputs and is an approximation otherwise.
  • domain assumption Straight-Through Estimator provides usable gradients through the binarizer
    Section 2.3 relies on STE for end-to-end training; the approximation is not validated against alternative gradient estimators.
  • domain assumption Concept annotations (SkinCon, WBCAtt) are reliable ground truth for the binarized concepts
    Sections 3.1 uses expert concept labels; if annotations are noisy or incomplete, the learned rules may be misleading.
  • domain assumption The annotated concept set is sufficient to explain the label
    The model is bottlenecked to the 48/24 annotated concepts; any label-relevant information outside the concept set cannot be captured by rules.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Concept-Driven Logical Rules for Interpretable and Generalizable Medical Image Classification." pith.science (2026). https://pith.science/paper/7WSSFPHQ

@misc{pith2026250514049,
  author       = {Pith},
  title        = {Pith review of: Learning Concept-Driven Logical Rules for Interpretable and Generalizable Medical Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7WSSFPHQ}},
  note         = {Machine review of arXiv:2505.14049}
}
read the original abstract

The pursuit of decision safety in clinical applications highlights the potential of concept-based methods in medical imaging. While these models offer active interpretability, they often suffer from concept leakages, where unintended information within soft concept representations undermines both interpretability and generalizability. Moreover, most concept-based models focus solely on local explanations (instance-level), neglecting the global decision logic (dataset-level). To address these limitations, we propose Concept Rule Learner (CRL), a novel framework to learn Boolean logical rules from binarized visual concepts. CRL employs logical layers to capture concept correlations and extract clinically meaningful rules, thereby providing both local and global interpretability. Experiments on two medical image classification tasks show that CRL achieves competitive performance with existing methods while significantly improving generalizability to out-of-distribution data. The code of our work is available at https://github.com/obiyoag/crl.

Figures

Figures reproduced from arXiv: 2505.14049 by the authors.

Figure 1
Figure 1. The architecture of Concept Rule Learner (CRL). CRL is a composition [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) An example of two adjacent logical layers, where the directed arrows [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. An illustration of concept logical rules for both skin disease diagnosis task [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 22 canonical work pages

  1. [1]

    Data in Brief30, 105474 (2020)

    Acevedo, A., Merino, A., Alférez, S., Ángel Molina, Boldú, L., Rodellar, J.: A dataset of microscopic peripheral blood cell images for development of automatic recognition systems. Data in Brief30, 105474 (2020)

    2020

  2. [2]

    In: International Conference on Machine Learning (2023)

    Barbiero, P., Ciravegna, G., Giannini, F., Zarlenga, M.E., Magister, L.C., Tonda, A., Lio, P., Precioso, F., Jamnik, M., Marra, G.: Interpretable neural-symbolic concept reasoning. In: International Conference on Machine Learning (2023)

    work page 2023

  3. [3]

    Elsevier (2017)

    Bolognia, J., Schaffer, J., Cerroni, L.: Dermatology. Elsevier (2017)

    work page 2017

  4. [4]

    Science advances 8(31) (2022)

    Daneshjou, R., Vodrahalli, K., Novoa, R.A., Jenkins, M., Liang, W., Rotemberg, V., Ko, J., Swetter, S.M., Bailey, E.E., Gevaert, O., Zou, J., et al.: Disparities in dermatology ai performance on a diverse, curated clinical image set. Science advances 8(31) (2022)

    work page 2022

  5. [5]

    In: Neural Information Processing Systems (2022)

    Daneshjou, R., Yuksekgonul, M., Cai, Z.R., Novoa, R.A., Zou, J.: SkinCon: A skin disease dataset densely annotated by domain experts for fine-grained debugging and analysis. In: Neural Information Processing Systems (2022)

    work page 2022

  6. [6]

    IEEE Transactions on Pattern Analysis and Machine Intelligence45(4), 4321–4334 (2023)

    Duan, X., Wang, X., Zhao, P., Shen, G., Zhu, W.: Deeplogic: Joint learning of neural perception and logical reasoning. IEEE Transactions on Pattern Analysis and Machine Intelligence45(4), 4321–4334 (2023)

    work page 2023

  7. [7]

    In: Ad- vances in Neural Information Processing Systems

    Espinosa Zarlenga, M., Barbiero, P., Ciravegna, G., Marra, G., Giannini, F., Dili- genti, M., Shams, Z., Precioso, F., Melacci, S., Weller, A., Lió, P., Jamnik, M.: Con- cept Embedding Models: Beyond the Accuracy-Explainability Trade-Off. In: Ad- vances in Neural Information Processing Systems. vol. 35, pp. 21400–21413 (2022)

    work page 2022

  8. [8]

    In: International Conference on Medical Image Computing and Computer Assisted Intervention (2024)

    Gao, Y., Gao, Z., Gao, X., Liu, Y., Wang, B., Zhuang, X.: Evidential concept em- bedding models: Towards reliable concept explanations for skin disease diagnosis. In: International Conference on Medical Image Computing and Computer Assisted Intervention (2024)

    work page 2024

Show all 28 references
  1. [9]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Groh, M., Harris, C., Soenksen, L., Lau, F., Han, R., Kim, A., Koochek, A., Badri, O.: Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fitzpatrick 17k Dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2021

  2. [10]

    In: Advances in Neural Information Processing Systems

    Hubara,I., Courbariaux,M., Soudry, D.,El-Yaniv,R., Bengio,Y.: Binarizedneural networks. In: Advances in Neural Information Processing Systems. vol. 29 (2016)

    2016

  3. [11]

    Nature Methods 18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: A self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203–211 (2021)

    2021

  4. [12]

    In: Proceedings of the 37th International Conference on Machine Learning

    Koh, P.W., Nguyen, T., Tang, Y.S., Mussmann, S., Pierson, E., Kim, B., Liang, P.: Concept Bottleneck Models. In: Proceedings of the 37th International Conference on Machine Learning. pp. 5338–5348 (2020)

    2020

  5. [13]

    arXiv preprint arXiv:2306.00890 (2023)

    Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890 (2023)

    2023 arXiv

  6. [14]

    Nature Communications15, 654 (2024)

    Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications15, 654 (2024)

    2024

  7. [15]

    arXiv preprint arXiv:2106.13314 (2021)

    Mahinpei, A., Clark, J., Lage, I., Doshi-Velez, F., Pan, W.: Promises and pitfalls of black-box concept learning models. arXiv preprint arXiv:2106.13314 (2021)

    2021 arXiv

  8. [16]

    Gao et al

    Margeloiu, A., Ashman, M., Bhatt, U., Chen, Y., Jamnik, M., Weller, A.: Do con- cept bottleneck models learn as intended? In: International Conference on Learning Representations workshop (2021) 10 Y. Gao et al

    2021

  9. [17]

    In: International Conference on Learning Representations (2023)

    Oikarinen, T., Das, S., Nguyen, L.M., Weng, T.W.: Label-Free Concept Bottleneck Models. In: International Conference on Learning Representations (2023)

    2023

  10. [18]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2024)

    Pang, W., Ke, X., Tsutsui, S., Wen, B.: Integrating clinical knowledge into concept bottleneck models. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2024)

    2024

  11. [19]

    arXiv preprint arXiv:1904.01554 (2019)

    Payani, A., Fekri, F.: Learning algorithms via neural logic networks. arXiv preprint arXiv:1904.01554 (2019)

    2019 arXiv

  12. [20]

    Nature Machine Intelligence1(5), 206–215 (2019)

    Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence1(5), 206–215 (2019)

    2019

  13. [21]

    arXiv preprint arXiv:2106.08641 (2021)

    Schrouff, J., Baur, S., Hou, S., Mincu, D., Loreaux, E., Blanes, R., Wexler, J., Karthikesalingam, A., Kim, B.: Best of both worlds: local and global explanations with human-understandable concepts. arXiv preprint arXiv:2106.08641 (2021)

    2021 arXiv

  14. [22]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Tsutsui, S., Pang, W., Wen, B.: Wbcatt: A white blood cell dataset annotated with detailed morphological attributes. In: Advances in Neural Information Processing Systems (NeurIPS). (2023)

    2023

  15. [23]

    Food and Drug Administration: Transparency for machine learning-enabled medical devices: Guiding principles (2024)

    U.S. Food and Drug Administration: Transparency for machine learning-enabled medical devices: Guiding principles (2024)

    2024

  16. [24]

    Medical Image Analysis 79, 102470 (2022)

    vanderVelden,B.H.M.,Kuijf,H.J.,Gilhuijs,K.G.A.,Viergever,M.A.:Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Medical Image Analysis 79, 102470 (2022)

    2022

  17. [25]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 46(02), 1121–1133 (2024)

    Wang, Z., Zhang, W., Liu, N., Wang, J.: Learning interpretable rules for scalable data representation and classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(02), 1121–1133 (2024)

    2024

  18. [26]

    In: International Conference on Learning Representations (2023)

    Yang, G., Song, L.: Learn to explain efficiently via neural logic inductive learning. In: International Conference on Learning Representations (2023)

    2023

  19. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yang, Y., Panagopoulou, A., Zhou, S., Jin, D., Callison-Burch, C., Yatskar, M.: Language in a Bottle: Language Model Guided Concept Bottlenecks for Inter- pretable Image Classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. ...

    2023

  20. [28]

    IEEE Transactions on Emerging Topics in Computational Intelligence 5(5), 726–742 (2021)

    Zhang, Y., Tiňo, P., Leonardis, A., Tang, K.: A Survey on Neural Network Inter- pretability. IEEE Transactions on Emerging Topics in Computational Intelligence 5(5), 726–742 (2021)

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.