List of Papers Browse by Subject Areas Author List
Abstract
Interpretable models are important for high-stakes medical image recognition, but often underperform unconstrained black-box architectures. We propose Question-Driven Decision Tree (QDT), a framework that formulates prediction as structured reasoning through sequential and interpretable sub-questions. To retain expressive power under such constraints, QDT adopts a teacher–student paradigm in which a high-capacity teacher distills knowledge to a logic-constrained student while preserving an auditable decision rule.Experiments on challenging dermatological benchmarks, including two public datasets (ISIC-2019 and MILK10k) and a private pressure injury dataset, show that QDT surpasses existing interpretable models and achieves performance comparable to, and in some cases exceeding, strong black-box methods, while maintaining auditable question-level decision paths.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1813_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{YaoXua_QuestionDriven_MICCAI2026,
author = { Yao, Xuancheng AND Zhang, Zhong AND Chen, Qiuhui AND Hong, Yi},
title = { { Question-Driven Decision Tree via Knowledge Distillation for Interpretable Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
1813 - MICCAI
The paper introduces QDT (Question-Drive Decision Tree), a novel framework for medical image analysis (classification) that is conceived to be interpretable by design. The authors use authoritative medical guidelines on the clinical task at hand to create a deterministic decision rule (M) that maps several binary sub-tasks to the final diagnosis; the sub tasks are sequential dichotomous questions whose answers are the leaves of the decision tree. For instance, in order to stage pressure injuries from images, the authors decompose the problem into five sub-tasks/questions: if the wound is obscured by slough, then the injury is unstageable; else, if the skin surface is intact, then. . . ; else, . . .
The architecture is as follows: a backbone is used to extract image features that are then processed by K+1 parallel branches, one for the main diagnostic task and K for each sub-task. While the “main” branch is a standard DL pipeline, each K-th binary problem is learnt by an MLP classifier, and then then predefined, deterministic rule M is used to combine all the predicted sub-task answer and give a final diagnosis; the “main” diagnosis and the deterministic one are then aligned during training to ground the final prediction in medical knowledge. Interestingly, in order to train all the sub-task classifiers, the dataset needs to contain annotations for all sub-questions for each image.
QDT naturally embeds structured reasoning thanks to the decomposition of the problem into interpretable and clinically accurate subsequent questions. According to the authors, however, the predefined, deterministic rule M constrains the expressive power of the model; to overcome this issue, they resort to knowledge distillation from a pre-trained QDT-based teacher. The teacher is free from these constraints because its sub-task predictions are combined via a learnable aggregator instead of the parameter-free deterministic rule.
The results suggest that QDT can provide accurate predictions grounded in medical knowledge. The authors test it on three dermatological datasets, and by using several CNNs and a ViT-B/16 as the backbone feature extractor (not sure about this, I talk about this in the additional comments). Also, QDT considerably outperforms interpretable-by-design models such as CBM (Koh, 2020) and DCR (Barbiero, 2023).
- 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.
QDT is an extremely interesting and, to the best of my knowledge, novel framework. It is based on a wise use of LLMs, i. e. , to extract binary questions from medical guidelines on the task at hand, and then rely on clinicians to validate the obtained diagnostic rule. The framework is well-structured, it outperforms two interpretable models and proves to improve performance on three CNNs and a ViT. Even considering the limitations, I think that this work is highly innovative and effective in bridging the accuracy-interpretability gap.
- 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.
While QDT allows the end-user to always know the predictions to the K sub-tasks related to the main problem, the framework still heavily relies on opaque DL models and attention-based feature fusion modules. So, can we really define QDT an “interpretable” framwork? As the authors state in the Conclusion, “assessing intermediate reasoning reliability and interpretability fidelity remains an open challenge”. Even if we do not consider QDT’s black-box components, another question arises: the sub-questions are clinically validated, but are they complete? Isn’t it possible that there are other questions that a clinician may be interested in to inspect and assess the prediction of the model? Maybe this was already considered by the authors and their collaborating clinicians, but it is not reported in the paper. Also, dataset annotations for QDT are not easy to obtain at all, as an expert is expected to give not only a final diagnosis, but also K additional sub-answers. Summing the three datasets involved in the study, the authors must have obtained such expensive annotations for more than 32,000 images.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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
1.I have spent some time trying to verify this, but I think Figure 2 does not match what is reported in “Attention-Based Feature Fusion” (2.1): 1.1 In the text, the authors talk about self-attention and gates for h_sub’s, but this is shown in the figure; 1.2 Fig. 2 shows that h_main goes through LayerNorm, then MLP, then sigma, whereas the text states that sigma(MLP(. )) processes both h_main and h_subs; 1.3 The FFN is said to be residual, but this is not shown in the figure; 1.4 The text states “FFN([h_main, h_cross]): does it mean that the FFN is fusing two different feature maps together? This is not shown in Fig. 2 and is not the standard cross-attention module, at least in the context of LLM or multi-modal models.
I understand that some of the points I raised may be due to the need for simplicity in the figures, but there are too many things that seem not to match between text and figure, and the reader would not know which source to trust.
2.The “Results” section show comparisons among three CNNs and a ViT, but the authors never state where exactly these architectures are used. In 2.1, the authors had previously stated, “We employ a convolutional neural network as the backbone of the feature encoder”. This aspect is very confusing and I think it needs to be clarified.
3.Similarly, no information is given about the K+1 encoders and MLP, apart from the fact that the former are “lightweight” (2.1).
4.Acronyms: the authors define PI (Pressure injury) three times (and do not use this acronym in “Results > Interpretability”) and KD (Knowledge distillation) twice.
5.In Introduction, the authors talk about “such strict logical constraints” to justify the use of KD, but at this point the reader (me, at least) does not really know what these constraints are.
6.In Experiments, “DCR” was never introduced. Also, the authors state that sub-task supervision acts as a “beneficial inductive bias” against sever class imbalance, but no evidence is provided for this claim. We do not know how the CNNs or the ViT would handle this class imbalance without QDT in the involved datasets.
- 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 focuses on DL interpretability in the medical domain in a highly original and effective way, and I think it should be accepted despite its limitations because it could be very influential for other researchers in the field.
- 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.
The authors addressed most of the points I raised in my comments and clarified them. Despite the reliance on an annotation system that is so hard to obtain and/or validate, I believe that this approach can be impactful for fellow researchers in the field, as it provides the end user with an auditable path (as the authors point out in the rebuttal) that could actually limit the opacity of decision making, at least at the latest stages of the model.
Review #2
- Please describe the contribution of the paper
A Question‑Driven Decision Tree framework is proposed that formulates medical image prediction as a sequential, clinically aligned reasoning process, while maintaining performance comparable to strong black‑box models. By combining a deterministic, interpretable logic tree with a teacher–student knowledge distillation strategy, the method successfully transfers complex visual decision boundaries from an unconstrained model to a fully auditable, logic‑constrained student.
- 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 use of sequential, hierarchy‑aware sub‑tasks goes beyond flat concept bottleneck models and better reflects how clinicians reason in practice. 2- The method is carefully engineered, combining multi‑task learning, attention‑based feature fusion, and multi‑level distillation in a coherent and principled way. 3- Empirical results are strong and consistent across multiple challenging dermatological datasets. 4- The teacher–student distillation strategy is well‑motivated and effective.
- 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 is no evaluation of how errors in individual sub‑tasks propagate through the decision tree, nor how stable the final prediction is to perturbations in sub‑task predictions. 2- The choice of a fixed number of sub‑tasks (e.g., five for pressure injury staging) is not theoretically or empirically justified, and the framework’s ability to handle variable numbers or alternative sub‑task structures is not explored. 3- Although interpretability is the primary motivation, the paper does not provide quantitative interpretability metrics nor a systematic qualitative comparison against existing interpretable approaches, limiting the strength of the interpretability claims. 4- The set of comparative baselines is limited and excludes several recent interpretable or neuro‑symbolic methods, which weakens the positioning of the proposed approach relative to the current state of the art.
- 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 does not provide sufficient information for reproducibility.
- 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?
The paper is well motivated and well engineered, however interpretability is positioned as the core contribution of the paper, yet it is not rigorously analyzed or validated.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Two of my main concerns about interpretability and comparative baselines are not addressed in the rebuttal. The first concern is more related as someone might be looking for an Interpretable method while looking into this paper as mentioned in the paper title ,”for Interpretable Prediction” but the paper content and the method has nothing to do with the interpretability , which might be a misleading signal for the readers.
Review #3
- Please describe the contribution of the paper
This paper proposes a Question-Driven Decision Tree (QDT) framework for interpretable medical image classification, which decomposes the final prediction into a series of clinically meaningful binary classification subproblems and alleviates the performance degradation caused by logical constraints as much as possible through teacher-student distillation. The authors conducted experiments on three skin-related datasets, namely Pressure Injury, ISIC-2019, and MILK10k, emphasizing that this method can achieve strong performance while maintaining auditable reasoning paths.
- 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 strengths of this paper lie in its intuitive overall idea and clear problem definition. Compared with ordinary black-box classifiers or flat concept bottlenecks, the authors attempt to formulate the prediction process as a structured decision-making process of “answering several consecutive questions”, which is appealing in terms of interpretability goals. The introduction of teacher-student distillation also completes the method chain: first, a high-capacity teacher is used to learn complex decision boundaries, and then knowledge is transferred to the student model constrained by a logical tree. In terms of experiments, the authors provide results, ablation studies, and certain interpretability demonstrations on three datasets, achieving acceptable overall completeness. Even though I am not an expert in this specific subfield, the main thread and core design of the paper are relatively easy to understand.
- 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 reservations are as follows: the most natural application scenario of this method seems to be tasks like pressure injury that inherently have relatively clear grading rules; when extended to ISIC-2019 and MILK10k, although the paper mentions that similar tree structures can be constructed based on dermatological heuristics such as the ABCDE rule, I am personally not fully convinced that these tree structures are equally sufficiently natural and unambiguous on these datasets. Secondly, from a methodological perspective, the paper is more like a relatively complete combination of multi-task framework, logical tree constraints, and knowledge distillation, which is reasonable overall, but the “core new methodological point” can be further refined. Finally, my research focus is more on natural images and general vision tasks, and I am not an expert in dermatology or medical interpretable reasoning. Therefore, my judgments on whether the design of these subproblems truly aligns with clinical practice and whether the coverage of relevant work is comprehensive enough are for reference only, and I kindly ask the AC to take this into account when weighing my review.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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 hereby issue a Weak Reject. This paper has a relatively clear problem definition and a complete method chain; the overall narrative of question-driven reasoning combined with knowledge distillation is valid, and the number of experiments is not insufficient. On the positive side, it does go further than general post-hoc explainability or flat concept expressions. However, I still have two main reservations at present: first, whether the naturalness and universality of such a question tree-based structure are strong enough across different tasks, especially when extended from pressure injury to other skin disease datasets; second, the method is more like a good integration solution rather than a particularly strong core methodological innovation. In addition, since I am not an expert in this field, my judgment will be relatively conservative. Overall, I believe this paper has a certain quality, but the current version is slightly below my acceptance threshold; if the rebuttal can more clearly explain the clinical basis and universality of the tree structure design, I will be willing to reconsider it.
- Reviewer confidence
Somewhat confident (2)
- [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
Author Feedback
We thank the reviewers and AC for their constructive comments and recognition of QDT’s motivation and potential. We address the following concerns. Q1:Tree design/fixed K [R2,AC] and completeness [R1] K is not tuned as a performance hyperparameter. It is induced by clinical taxonomy and dataset sources for auxiliary labels: K=5 for pressure injury (PI) following NPIAP, and K=7/8 for ISIC-2019/MILK10k using ABCDE-inspired heuristics and dataset structure. Two melanoma datasets use different trees, so QDT has no global K/topology. Tree construction: guidelines/dataset definitions are organized into binary questions and validated by clinicians for correctness and completeness w.r.t. the target rule. LLM only structures clinical cues, not inventing criteria. Questions are extensible. We will clarify that QDT is flexible, not universally optimal. Q2:Generalizability/naturalness [R3,AC] We agree with R3 that PI staging is the most natural case because its rules are explicit. Dermatology datasets are less canonical and may admit multiple decompositions. Thus, QDT generalizes at the framework level, not through a universal tree. It applies when diagnosis can be decomposed into clinically meaningful questions from criteria or heuristics. We will revise this scope and avoid implying a universally optimal tree. Q3:Annotation requirements [R1,AC] QDT uses intermediate supervision but does not require experts to manually provide K new answers for every image. For PI, answers follow NPIAP staging criteria. For ISIC-2019 and MILK10k, labels are constructed from ABCDE-aligned descriptors, metadata, and clinical heuristics, with clinician validation. Metadata derive auxiliary labels only, not model inputs. The main effort is validating taxonomy and label rules, not exhaustive re-annotation. No intermediate annotations or metadata are needed at test time. Q4:Interpretability, baselines, and positioning [R1-R3,AC] QDT does not make DL backbones or attention modules transparent. Its interpretability is at question and decision levels: each Q_i is clinically meaningful, clinician-validated, and exposed, and M_S maps predicted sub-answers to the final diagnosis, yielding an auditable path, not just a black-box probability. The student’s final prediction is determined by M_S; the main head, fusion, and KD improve representations but do not replace M_S. QDT differs from post-hoc methods by exposing a reasoning path, and from CBM by using an explicit clinical aggregation rule; CBM/DCR are representative baselines. The technical core is teacher-student asymmetry, not generic MTL+tree+KD: a soft aggregator trains the teacher, while the student uses the deterministic rule. We will temper broad claims, expand related work, and discuss fidelity/neuro-symbolic comparisons as limitations. Q5:Error propagation/stability [R2] Since M_S is deterministic, error propagation can be analyzed structurally. For path P with independent question errors p_i, correct traversal has probability \prod_{i\in P}(1-p_i), and path error is bounded by 1-\prod_{i\in P}(1-p_i)\approx\sum_{i\in P}p_i for small p_i. Thus, root errors may matter more. Under hard-thresholded inference, off-path perturbations do not affect the current decision; only on-path branch changes alter the output. QDT mitigates this via sub-task supervision, KD alignment, and clinical tree construction. We will add this analysis and note perturbation-based stability as future work. Q6:Presentation/reproducibility [R1,R2] We will align Fig. 2 with Sec. 2.1 by showing gated self-attention, FFN residual, and FFN input [h_main,h_cross]. We will use “backbone (CNN or Transformer)”, clarify shared backbone/lightweight heads instead of K+1 full encoders, define PI/KD/DCR, tone down unsupported class-imbalance claims, and add head/fusion details. Upon acceptance, code/configs, Q_i definitions, and prompts will be released. For private PI data, we will document construction, labels, and governance/de-identification.
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 paper proposes a Question-Driven Decision Tree (QDT) framework, which decomposes the medical image classification task into sequential, clinically meaningful binary classification sub-tasks. The framework is argued to be interpretable by design despite relying on DL backbones. There are serious concerns raised by reviewers, including the choice of a fixed number of sub‑tasks, heavy annotation requirements, and the generalizability of the framework. Authors are encouraged to address all the reviewers’ concerns.
- 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 clarified design and generalization-related issues raised by the reviewers. Still, there are concerns about limited interpretability stemming from DL-based feature extractors and the lack of comparisons with recent neuro‑symbolic methods. In my opinion, the work is a right step toward responsible, deployable medical AI models. Moreover, the training of such a tree-based framework is not trivial. The paper tackles this problem with an elegant KD trick, which is interesting. I feel the paper could be of interest to the community.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
Despite the clinical motivation and interesting concept of a Question-Driven Decision Tree (QDT) framework, the paper is recommended for rejection due to fundamental gaps in validating its core claims.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The paper introduces an interesting approach for achieving interpretability in medical image diagnosis. The reviewers have recognized the importance of this work and appreciated the neat design of the framework.
However, several valid concerns were raised by the reviewers in the first round. The authors have tried to argue convincingly for some of them. However, I agree with the R#2 that the major claim about interpretability should be supported by metrics to measure it. Merely introducing a decision-tree based question-answering approach doesn’t make it interpretable although the combination teacher-student distillation strategy and the decision-tree based hierarchy are interesting additions that go beyond the current concept bottleneck models.
Despite the above two interesting design choices, the main issues lie with the systematic evaluation of how this interpretability is measured through a metric rather than just the introduction of the decision-tree. The other point raised by R#2 is also valid that the decision-tree based approach is also similar to the neurosymbolic frameworks which have logical rules to determine the final prediction. Such approaches are not explored or compared in the paper.
I think overall, it’s an interesting paper and approach, however, given the current two major limitations noted above, I think it’s not above the acceptance level. The authors should consider the comments positively to improve their work for their future submissions.