List of Papers Browse by Subject Areas Author List
Abstract
Large-scale pre-trained models have achieved remarkable performance in CT organ segmentation, yet individual models exhibit complementary strengths across different anatomical regions. While multi-teacher knowledge distillation can leverage these complementary advantages, existing methods assign teacher weights at the instance or global level, overlooking the inherent spatial heterogeneity of CT volumes where tissue contrast, boundary ambiguity, and organ morphology vary significantly across regions. We propose Task-Performance Routing (TPR), a novel framework that dynamically routes to the best-performing teacher at each spatial region based on actual prediction errors rather than learned parameters. TPR employs hierarchical spatial adaptation—fine-grained routing in shallow layers for boundary details and coarse-grained routing in deep layers for semantic coherence—and applies performance-based selection in hard regions while encouraging complementary learning in easy regions. Extensive experiments on BTCV, MSD-Liver, and MSD-Pancreas datasets demonstrate that TPR consistently outperforms individual teachers, adaptive ensemble methods, and existing multi-teacher distillation baselines, achieving superior segmentation accuracy without inference overhead. Code is available at \url{https://github.com/EdvinCecilia/TPR_MultiTeacher}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0973_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/EdvinCecilia/TPR_MultiTeacher
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanJia_TaskPerformance_MICCAI2026,
author = { Yang, Jiaye AND Liu, Yao AND Dai, Die AND Zhang, Hanwen AND Jiang, Peiyuan AND Liu, Qiao AND Xie, Yutong AND Wang, Peng},
title = { { Task-Performance Routing for Multi-teacher Distillation in CT Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a multi-teacher knowledge distillation method. It dynamically selects the optimal teacher model for each spatial region, achieving superior segmentation accuracy without introducing additional inference overhead.
- 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 framework adopts a hierarchical fusion design, which selects the best-performing teacher for hard regions while encouraging complementary fusion in easy regions. 2.Evaluating their performance across multiple datasets prove the generalizability of the model.
- 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.Figure 1 only presents the distinct spatial activation patterns of the two teacher networks, but does not demonstrate the complementary spatial activation patterns learned by the student network. 2.Figure 2 describes that the Routing Module Mechanism requires computing both region-based error and confidence. However, only region-wise cross-entropy errors are introduced in the manuscript, with no mention of confidence. 3.In the Routing Weight Generation section, it is stated that region-wise cross-entropy errors are computed from the final logits and the ground truth. However, the description only explains how to calculate region-wise cross-entropy errors based on the final layer, without mentioning how the l-th layer is processed. 4.The caption of Table 1 states that the unit of Dice is %, yet the Dice value of 0.8545 presented in the table is unreasonably low. Furthermore, it can be observed that the performance improvement of the TPR in multi-organ segmentation is not significant. 5.The proposed method involves an excessive number of hyperparameters, which raises concerns about the robustness of the TPR. 6.The ablation study in Table 4 appears incomplete. 7.Tables 2, 3, and 4 present ablation studies on the MSD-Pancreas dataset, while Figure 3 shows visualization results on the BTCV dataset, leading to inconsistency in content. In addition, there is no comparison with other methods in Figure 3, making it impossible to demonstrate the advantages of the TPR.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This manuscript presents an innovative multi-teacher segmentation approach. Nevertheless, significant revisions are needed to resolve problems including ambiguous expression and inadequate ablation experiments, which are essential to validate the effectiveness and rationality of the proposed method.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the author feedback, I recommend acceptance. The authors have adequately addressed my main concerns.
Review #2
- Please describe the contribution of the paper
This paper presents Task-Performance Routing (TPR), a multi-teacher distillation framework for 3D CT organ segmentation. The method assigns teacher weights region-wise according to segmentation error, uses a hierarchical spatial routing scheme, and applies advantage-based teacher selection in hard regions while encouraging uniform fusion in easy regions. Experiments on BTCV, MSD-Liver, and MSD-Pancreas, with two teacher pairs, show consistent improvements over individual teachers, adaptive ensembling, and several multi-teacher knowledge distillation baselines, without extra inference cost.
- 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 main idea of routing teachers based on region-wise performance is simple, intuitive, and well motivated. 2.The split between hard and easy regions is a reasonable design that helps avoid over-committing to a single teacher in easy areas. 3.The evaluation covers three datasets and two teacher pairs, which provides broad evidence of effectiveness. 4.The ablation study is informative and separates the contributions of global weighting, spatial routing, hierarchical granularity, and loss terms.
- 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.The routing relies on ground-truth labels to compute region-wise errors during training. This makes the method less applicable. Although the paper acknowledges this point, it remains an important limitation. 2.The figures in the paper appear blurry, and some text is difficult to read. 3.In the codebase, hard_region_mask seems to be used only for logging, and there is no clear implementation of a separate hard/easy region mechanism. 4.The reported gains are relatively modest in several settings, but the paper does not provide enough information about variance, confidence intervals, or multi-seed stability. This makes it harder to judge the robustness of the improvements. 5.The advantage-based weighting can still assign positive weights to a teacher even when that teacher is worse than the current student in a region. 6.The method is formulated in a way that could extend beyond two teachers, but experiments only consider the case of two teachers. It is therefore unclear how the approach scales when the number of teachers increases, especially for routing stability and balance.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper has a clear and intuitive idea, and the experiments show consistent gains with useful ablations. However, the method relies on ground-truth-based routing during training, some implementation details are unclear, and robustness and scalability are not sufficiently validated.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
I maintain Weak Accept after the rebuttal. The authors clarify several presentation issues, and the core idea, ablations, and multi-dataset evidence remain solid. Three concerns remain: (1) My W5 is not addressed: advantage-based softmax can still assign positive weight to teachers locally worse than the student, and the rebuttal does not engage with this soundness question. (2) Robustness of Table 1 (W4 / AC) is mis-answered. Paired t-tests on ablations establish the significance of individual components, not multi-seed variance on the headline numbers, and several reported gains are modest. (3) The code-paper discrepancy (W3 / AC) is acknowledged but resolved only by promise, which cannot be verified at rebuttal time. Since the hard/easy switch is a central claim, this weakens reproducibility.
Review #3
- Please describe the contribution of the paper
The paper tackles the challenge of Multi-Teacher Knowledge Distillation (MTKD) in 3D CT organ segmentation. The authors observe that different large-scale pre-trained foundation models (e.g., SuPreM, VoCo, CLIP-Driven) exhibit complementary strengths in different spatial regions and organs due to their distinct pre-training paradigms. To effectively fuse this knowledge, the authors propose Task-Performance Routing (TPR). Unlike existing methods that use global weighting or learnable parameters, TPR dynamically assigns teacher weights at a region-wise level based on actual prediction errors. A core design is the hierarchical spatial adaptation: it applies fine-grained 3D grid routing in shallow layers to capture local boundary details, and coarse-grained routing in deep layers to maintain global semantic coherence. Furthermore, it explicitly selects the best teacher in “hard” regions while applying uniform weighting in “easy” regions to encourage complementary learning.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Elegant Handling of 3D Spatial Heterogeneity: The paper presents a highly intuitive and medically relevant premise—that tissue contrast, boundary ambiguity, and organ morphology vary drastically across a single 3D CT volume. The hierarchical 3D region partitioning aligns with the spatial abstraction levels of modern CNN/Transformer encoders.
Parameter-Free Routing Mechanism: Relying on the cross-entropy error to route to the best teacher is a clever, parameter-free solution. This avoids the need for complex, learnable routing networks (like those in Mixture-of-Experts) which are prone to overfitting on the limited data typically available in medical imaging downstream tasks.
Use of Strong, Modern Baselines: The authors benchmark their method using contemporary, state-of-the-art medical foundation models (SuPreM, VoCo, CLIP-Driven) and modern architectures (SwinUNETR). This makes the results highly relevant to the current landscape of medical image analysis.
Extensive Ablation Studies: The experimental validation is rigorous. The component-wise ablations (Table 2), granularity strategy comparisons (Table 3), and loss component analyses (Table 4) isolate and mathematically prove the contribution of every single design choice in the framework.
- 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.
Dependence on Ground Truth for Routing: A significant limitation (which the authors briefly acknowledge) is that the routing mechanism fundamentally relies on calculating the prediction error against the ground truth during training. This strictly limits TPR to fully supervised settings, preventing its application in semi-supervised or unsupervised domain adaptation scenarios where MTKD is traditionally highly valuable.
Limited Dataset Scale for Foundation Model Distillation: While the teachers are foundation models pre-trained on massive datasets, the downstream evaluation is performed on relatively standard, small-scale datasets (BTCV, MSD-Liver, MSD-Pancreas). Evaluating the distillation framework on a larger, more complex benchmark like AMOS22 or AbdomenCT-1K would make the claims of “superior knowledge fusion” much more robust.
High Training Memory Overhead: Forward-passing multiple large models (e.g., two SwinUNETR teachers and one student) simultaneously to compute features and logits requires immense GPU memory. The paper would benefit from a brief discussion on the computational complexity and memory footprint during the training phase, as this limits the batch size and the scalability to N > 2 teachers.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The Task-Performance Routing framework is an elegant, parameter-free solution to a highly relevant modern problem: how to optimally fuse the diverging strengths of various medical foundation models. The hierarchical grid routing demonstrates a strong understanding of deep learning feature abstraction, and the ablation studies are thorough in justifying the architecture. While the absolute dependence on ground truth limits the method’s extension into semi-supervised learning, and the training memory overhead is likely substantial, the contribution to fully supervised multi-teacher distillation is significant. The methodological soundness, combined with empirical validation against modern baselines, makes this an interesting paper to the community.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 has clarified my concerns to some extent. I am in favor of a borderline acceptance.
Author Feedback
We thank the reviewers and AC for their constructive feedback and for recognizing TPR’s motivation, parameter-free hierarchical hard/easy routing, multi-dataset evaluation, and inference-free deployment. We address the main concerns below.
(1) AC/R1 - Intermediate-layer routing. Sec. 2.2 was unclear on how final-logit errors are used for intermediate layers. TPR computes voxel-wise cross-entropy error maps from the final logits of the student and each frozen teacher, then pools these maps with the stage-specific 3D grid to obtain region-level errors for each encoder stage. Thus, shallow layers use finer grids for local boundary ambiguity, while deeper layers use coarser grids for semantic coherence. We will clarify this output-space error pooling and stage-wise routing in Sec. 2.2 and Fig. 2. (2) AC/R1 - Confidence vs. error. The finalized TPR formulation is error-based and does not use confidence as an additional routing signal. The word “confidence” came from an exploratory draft of Fig. 2 and caused confusion. We will remove this wording and consistently describe the router as task-error-driven routing.
(3) AC/R1/R2 - Robustness, ablations, and notation. The submitted manuscript already reports paired t-test significance for the ablation studies in Tables 2-4 (p<0.05), showing that the main routing, granularity, and loss designs contribute consistently. Table 1 also shows that TPR improves over individual teachers, adaptive ensemble, and MTKD baselines across three datasets and two teacher-pair settings, rather than in a single isolated case. We will make this existing evidence clearer in the text. We will also clarify the loss-component ablation in Table 4 and correct the Dice notation in Table 1: values such as 0.8545 are Dice fractions and should be reported as 85.45% or clearly labeled as fractions.
(4) R1/R2 - Figure clarity and qualitative comparison. We will improve the resolution and readability of all figures. Fig. 1 is intended as motivation, showing teacher complementarity, while Fig. 3 already includes the student prediction and Student Advantage map. We will revise the captions and cross-reference Fig. 3 more clearly to show how the student benefits from teacher complementarity, and clarify the teacher/student comparisons.
(5) AC/R2/R3 - Scope, scalability, and training memory. TPR targets fully supervised downstream distillation: labels are needed only during training to estimate region-wise task errors, and the deployed model is a single student. In semi-supervised or label-scarce settings, unreliable or missing labels would make error-based routing less reliable; we will clarify this limitation and leave label-free or pseudo-label-based routing as future work. Our experiments focus on two teachers and three public CT benchmarks covering multi-organ, liver, and pancreas segmentation; evaluation on larger teacher pools and larger benchmarks such as AMOS/AbdomenCT-1K remains future work. The routing rule itself is not restricted to two teachers: hard-region weights can be computed over all teacher advantages, and easy-region fusion can assign equal weights to all teachers. Regarding cost, teachers are frozen and evaluated under no_grad, so no teacher backward graph is stored; sequential teacher forward can reduce peak memory. TPR has no inference-time overhead.
(6) AC/R2 - Code-paper discrepancy. The initially released code package mistakenly contained an earlier testing branch where hard_region_mask was mainly used for monitoring/debugging, causing the impression that the hard/easy mechanism was not implemented. The finalized implementation used for the intended TPR mechanism explicitly uses hard_region_mask to switch routing weights: hard regions use advantage-based softmax routing, while easy regions use uniform complementary fusion, following Eq. (2). We will ensure the code released after acceptance matches this implementation.
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.
Scores are 3/4/4.The two weak accepts emphasize methodological soundness; the weak reject (R1, faculty) is driven by ambiguous exposition and incomplete ablations rather than disagreement with the core idea. The concerns are largely addressable, so the paper is invited for rebuttal. Authors should prioritize (a) clarifying the intermediate-layer routing formulation and the role of confidence vs. error in the routing module, (b) reporting multi-seed variance on the headline numbers, (c) discussing scalability to N>2 teachers and training-time memory footprint, and (d) reconciling the code-vs-paper discrepancy around the hard/easy region mechanism.
- 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.
Reviewers viewed the paper favorably overall, highlighting Task-Performance Routing (TPR) as a simple, intuitive, and well-motivated multi-teacher distillation framework that exploits the complementary strengths of different medical foundation models through region-wise routing. Strengths included the parameter-free routing mechanism, hierarchical spatial adaptation across feature levels, strong empirical performance on multiple CT segmentation benchmarks, comparisons with modern foundation-model teachers, and comprehensive ablation studies validating key design choices. Initial concerns focused on reliance on ground-truth errors for routing during training, limited evaluation to two-teacher settings, unclear implementation details and presentation issues, modest performance gains in some settings without multi-seed variance analysis, training-memory overhead, and questions regarding scalability and robustness. These concerns led to one Weak Reject and two Weak Accept recommendations before rebuttal. Following the authors’ responses, reviewers agreed that the major methodological and presentation concerns had been sufficiently clarified, resulting in all reviewers supporting acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After the rebuttal, all reviewers agreed that the paper should be accepted. The Area Chair supports this consensus and recommends acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All reviewers agree to accept the paper. I do agree with R2 about the concerns about result robustness and more ablation results. However, this leads to a greay area in the final decision as I’m not sure whether this may requires additional experiments which against MICCAI’s rebuttal rules. Here I recommend a poster presentation of this work, not oral. The authors should also try to correspond to R2W5 with a few words (no need to be detailed as long as the improvements introduced in this paper are proved to be statistically significant).