List of Papers Browse by Subject Areas Author List
Abstract
Epidermal growth factor receptor (EGFR) mutation prediction and progression-free survival (PFS) prediction are critical tasks for targeted therapy planning in lung cancer and can be jointly modeled from the same chest CT scan. In real-world clinical datasets, supervision is often inconsistent across patients: not all patients have labels for both tasks and tumor region annotations, leading to heterogeneous supervision across different patient subsets. However, most existing multi-task learning methods leverage a single shared representation across tasks, with limited consideration of representation consistency during training on these differently supervised patient subsets. In this study, we propose the Tumor-Guided Bridging Distillation Network (TGBD-Net), a unified framework for stabilizing representation learning under heterogeneous supervision, enabling joint EGFR mutation and PFS prediction. The proposed Tumor-Guided Feature Extraction (TGF) module regularizes feature learning using available tumor region annotations, while maintaining consistent behavior when such annotations are unavailable. In addition, a Bridging Distillation (BD) module enables cross-task knowledge transfer, allowing patients with single-task labels to contribute to joint representation learning. Experiments on a cohort of 7,348 patients with lung cancer demonstrate that TGBD-Net achieves an AUC of 0.823 for EGFR mutation prediction and a C-index of 0.743 for PFS prediction, highlighting its effectiveness in real clinical settings. Our code is available at https://github.com/MenchYoung/TGBD-Net.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0569_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/MenchYoung/TGBD-Net.git
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanHui_TGBDNet_MICCAI2026,
author = { Yang, Huihui AND Wang, Shuo AND Huang, Xingyu AND Sang, Haolin AND Liu, Chengcai AND Wu, Yi AND Tian, Jie},
title = { { TGBD-Net: Tumor-Guided Bridging Distillation for Joint EGFR Mutation and Survival Prediction in Lung Cancer from CT Imaging } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a Tumor-Guided Bridging Distillation Network (TGBD-Net), a unified framework designed to enable stable feature representation learning under heterogeneous supervision, while simultaneously supporting EGFR mutation prediction and PFS estimation.
- 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 main contributions are summarized as follows: (1) a unified framework that incorporates patients with single-task labels into joint learning; (2) two complementary representation-level mechanisms that regularize feature formation without requiring tumor annotations for all patients, and enable knowledge transfer across patient subsets with different task labels in a shared representation space; (3) extensive experiments on 7,348 patients demonstrating the robustness of TGBD-Net across multiple tasks.
- 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.
Methodology: 1、The organization and input pipeline of data for different tasks are not clearly described (i. e. , how different tasks are constructed). For each task, what do the classification and regression heads specifically output? 2、 How is qtg obtained and trained? 3、 Compared with other cross-attention mechanisms, what are the key innovations of the TGF module? 4、 How are the teacher and student models explicitly defined in the bridging distillation process? Experiments: 1、 The authors do not explain why TGBD-Net does not achieve the best performance in terms of sensitivity and specificity. 2、Although extensive comparative experiments are conducted, the paper does not present the detection performance of single-task learning under different supervision settings.
- 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 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.
(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 authors present a notable and innovative idea that effectively addresses a challenging problem in this domain.
- Reviewer confidence
Very confident (4)
- [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
Review #2
- Please describe the contribution of the paper
This paper proposes TGBD-Net, a multi-task learning framework for joint EGFR mutation status classification and progression-free survival (PFS) prediction from chest CT scans . The key idea is that real-world clinical datasets include patients with varying label availability and partial tumor-region annotations. Two complementary modules are proposed: (1) a Tumor-Guided Feature Extraction (TGF) module that uses cross-attention to incorporate tumor region information when bounding box annotations are available, and (2) a Bridging Distillation (BD) module that transfers cross-task supervision through a set of periodically updated anchor prototypes, enabling partially labeled patients to contribute to joint representation 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.
1.Clinically motivated problem formulation. Heterogeneous supervision (partial labels, partial annotations) is a genuine, under-addressed challenge in real-world clinical ML. Framing MTL under this constraint is practically relevant. 2.Well-designed BD module. The anchor-based bidirectional distillation mechanism is technically interesting; using K-means prototypes with Hungarian matching to preserve identity across updates is a thoughtful design choice. The two-stage (attribute update → soft label distillation) formulation is clean and clearly explained. 3.Large-scale evaluation. 7,348 patients are a large cohort for this task domain, lending credibility to the reported performance figures. 4.Thorough ablations. The three ablation studies — module combinations, alignment strategies/feature choices, and anchor parameterization strategies — are well-structured and provide meaningful insight into each design decision.
- 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.Single-institution data, no external validation. All 7,348 patients are from one hospital. Without external validation, it is unclear whether the learned representations generalize across scanners, protocols, or institutions. 2.Severe label imbalance in the joint set. Only 383 patients have both EGFR and PFS labels, versus 6,484 with EGFR-only. The BD module is designed to exploit partially labeled samples, but with such an extreme imbalance, the learned anchor attributes for survival may be unreliable. The paper does not analyze sensitivity to this imbalance. 3.PFS test set is very small. The test set contains 735 patients total, but only a fraction have PFS labels (roughly ~96 based on the 8:1:1 split of 481 PFS patients). KM curves show only 86 patients. 4.TGF module design has limited novelty. The tumor-guided cross-attention with conditional alignment is a reasonable engineering choice, but cross-attention with partial supervision is well-established. The claim of novelty here should be more carefully scoped relative to existing tumor-region attention methods in the CT radiogenomics literature (e.g., the MICCAI 2025 paper by Huang et al. (Feature copy paste) on EGFR prediction is cited but not compared against). 5.No comparison against the most relevant prior work on heterogeneous MTL for medical imaging. JTR-MTL (Nishi et al., CVPR 2024) is the closest baseline and was designed specifically for partially labeled MTL, yet the paper does not discuss why TGBD-Net’s BD mechanism is preferable; the analysis remains quantitative only. 6.Hyperparameter sensitivity not analyzed. K=4 anchors, τ=0.2, α=0.9, β=0.5 are set without any sensitivity analysis. Given that these govern the distillation quality, this is an important omission. 7.Model selection criterion is asymmetric. The best model is chosen based on validation AUC (EGFR), and the same checkpoint is used for PFS evaluation. This may systematically favor EGFR performance; a joint validation criterion would be more principled.
- 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.
(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 addresses a real clinical problem, and the BD mechanism is technically novel. If the authors address the weakness, I can think about improving the score.
- Reviewer confidence
Very confident (4)
- [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
Review #3
- Please describe the contribution of the paper
This paper proposes TGBD-Net for joint EGFR mutation and PFS prediction from chest CT with heterogeneous supervision. I think the main contribution is that the authors combine anchor-based task memory, task-specific attribute accumulation, and similarity-weighted soft label distillation in a quite nice way. Because of this design, cases with only one task label can still provide some information for joint training. The tumor-guided feature module is another helpful part, since it uses available tumor annotations to guide the shared representation, but does not require annotation for every patient. Overall, the contribution is mainly this combined framework for partially labeled multi-task 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.
A main strength of the paper is the BD design. I think the way the authors connect anchor-based task memory, task-specific attribute accumulation, and similarity-weighted soft label distillation is quite nice. This makes the framework more than a standard shared-backbone multi-task model, and it gives a reasonable way for partially labeled cases to still contribute across tasks through the shared anchors.
Another strength is that this design may help the model capture EGFR-associated radiographic patterns in a more structured way, instead of only relying on a single global shared feature. Together with the tumor-guided feature module, the method tries to organize phenotype-related information in the shared representation while still handling missing labels and missing tumor annotations.
- 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.
The survival part is still a weakness. PFS is one of the main endpoints, but the paper does not clearly explain the PFS definition, censoring details, or treatment background. Also, the number of PFS cases is much smaller than the EGFR cohort, so the survival results need stronger support.
Another weakness is that some interpretation is a bit too strong. TGF alone improves EGFR prediction, but it does not improve the C-index over the baseline, and the tumor-guided feature is much worse than the final fused feature for PFS. So while the full model works better overall, the role of TGF should be discussed more carefully.
- 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?
I gave this paper a weak accept because I think the idea is interesting and the overall method is reasonably well designed. In particular, the BD part is a nice way to connect partially labeled cases across tasks, and the full model gives the best overall results in both EGFR prediction and PFS prediction. The ablation study also supports that the full design is better than the simpler variants. I still have some concerns about the survival part and a few claims that could be written more carefully, but overall I think the paper has enough technical value and practical relevance to be accepted.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
1.TGF novelty and methodological role Re to R1-C3, R2-C4: The key contribution of TGF is enabling tumor-guided representation regularization under partially available annotations, where attention patterns learned from annotated samples can be implicitly transferred to unannotated patients. Our work primarily focuses on joint EGFR/PFS learning with partially available labels and annotations, which differs from FCP-Net (MICCAI 2025 Huang et al.), which addresses slice-thickness variability through feature copy-paste for EGFR prediction. Re to R3-C2: TGF mainly enhances tumor-focused features, which is more beneficial for EGFR prediction due to tumor morphological patterns in CT, while PFS prediction may require broader surrounding tissue environment. Therefore, the full TGF+BD framework achieves the best overall PFS performance. 2.BD Module and difference from existing MTL methods Re to R1-C4: In BD, anchor-level attributes accumulated from labeled patients serve as teacher signals, while partially labeled patients act as students through bidirectional soft distillation (Sec. 2.2). Re to R2-C5: Compared with JTR-MTL and other optimization-level MTL approaches that mainly address task conflicts or dynamic weighting, BD performs representation-level regularization through bridging anchors in a shared semantic space. 3.Survival cohort, heterogeneous supervision and generalization Re to R2-C1/3, R3-C1: PFS was defined as the time from the initiation of treatment (EGFR-TKI) to the progression of the disease with regular follow-up; 18.5% cases were censored. External multicenter validation will be explored in future work. The PFS cohort is smaller than the EGFR cohort due to the limited availability of longitudinal follow-up data in real-world clinical settings. Nevertheless, the cohort includes heterogeneous acquisition conditions (e.g., varying CT slice thicknesses of 3–8 mm), partially mitigating overfitting to a single acquisition condition. Additionally, compared with many imaging-based survival studies, the current cohort remains relatively large-scale for PFS modeling. Re to R2-C2: To alleviate label imbalance, we adopted a weighted sampling strategy to increase the exposure of PFS samples during training (Sec.3.2). The motivation of BD is to improve PFS learning through joint supervision from the larger EGFR cohort under heterogeneous labeling. Despite limited PFS labels, TGBD-Net achieved the best survival performance among all baselines, supporting the effectiveness of the learned anchor attributes. 4.Experiments and evaluation Re to R1-Exp1/2: Regarding sensitivity/specificity, TGBD-Net achieves the best overall balance between the two metrics while also obtaining the best AUC and C-index. Table 1 includes corresponding single-task baselines (e.g., ResNet, Swin Transformer, DeepSurv), while Sup.-MTL demonstrates that simple joint learning improves over separate single-task modeling. Heterogeneous supervision is only defined in the joint multi-task setting. Re to R2-C7: Since the EGFR cohort is larger, validation AUC provides a relatively more stable selection criterion. All methods were evaluated under the same selection protocol to ensure fair comparison, and TGBD-Net still achieved the best survival performance, suggesting that the learned shared representation does not overfit to EGFR alone. 5.Additional clarifications Re to R1-C1/2: The dataset can be divided according to task labels (EGFR-only/PFS-only/both) and annotation availability, which are unified through the loss routing in Eq. (5). The EGFR head outputs classification logits, while the PFS head outputs a survival risk score. qtg is a learnable token with the same feature dimension as fg in TGF, acting as a tumor-guided prior optimized end-to-end (Sec 2.1). Re to R2-C6: Hyperparameter sensitivity experiments were conducted for K=2/3/4/6, α=0.8/0.9, β=0.5/1.0.The best-performing settings were selected. Due to page limitations, complete results were omitted.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
All positive reviews
