List of Papers Browse by Subject Areas Author List
Abstract
Glioma grading and molecular subtyping are challenging due to significant intratumoral heterogeneity and inter-task interference in multi-task learning settings. Existing multi-task learning approaches often fail to effectively disentangle shared and task-specific representations, limiting their ability to capture complex biological patterns. To address these issues, we propose the Distillation Expert Collaboration Network (DEC-Net), which comprises two key modules. First, a Distillation ResModule utilizes a knowledge distillation mechanism to guide a ResNet backbone in learning discriminative multimodal representations from tumor-relevant regions. Second, a Progressive Mixture-of-Experts module hierarchically integrates shared and task-specific experts, enabling layer-wise adaptation of representation spaces to mitigate inter-task interference. Furthermore, a multi-task dependency loss is introduced to explicitly model the biological correlation between glioma grades and subtypes, thereby improving prediction consistency. Extensive experiments conducted on a large hybrid cohort comprising UCSF-PDGM and EGD datasets, along with external validation on the TCGA dataset, demonstrate that DEC-Net achieves superior performance and robust generalization compared with state-of-the-art methods for non-invasive glioma diagnosis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4563_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{LiuFei_DECNet_MICCAI2026,
author = { Liu, Fei AND Cheng, Jianhong AND Tian, Xu AND Wang, Jianxin AND Hu, Yue AND Liu, Jin AND Li, Hong-Dong},
title = { { DEC-Net: A Multi-Task Distillation Expert Collaboration Network for Glioma Grading and Molecular Subtyping Using Multimodal MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes DEC‑Net, a Distillation Expert Collaboration Network for joint glioma grading and molecular subtyping from multimodal MRI. The framework integrates three complementary components: (1) a Distillation ResModule (DRModule) that employs an anomaly‑guided teacher-student mechanism to focus learning on tumor‑relevant regions; (2) a Progressive Mixture‑of‑Experts (PMoE) module that balances shared and task‑specific representation learning to mitigate inter‑task interference; and (3) a Dependency Loss (D‑Loss) that explicitly models biological correlations between grade, IDH mutation, and 1p/19q status. Extensive experiments on multiple datasets (UCSF‑PDGM, EGD, and TCGA) show improved AUCs and accuracy over previous CNN‑based multi‑task baselines, with consistent external‑test performance demonstrating good generalizability.
The integration of anomaly‑guided distillation and biologically informed dependency modeling is interesting and moderately novel. While the DRModule and PMoE build on existing ideas, the D‑Loss provides a distinctive, biologically meaningful regularization contribution.
- 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 paper is clearly written, logically organized, and experimentally thorough. The authors evaluate on both internal and external cohorts and provide ablation studies confirming the contribution of each module. The method’s performance remains strong under domain shift, suggesting good robustness.
- 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.
Some methodological details could be clarified, for example, the rationale for center cropping (which may exclude boundary information) and whether inference used patch aggregation or whole‑volume prediction. The paper does not discuss its own limitations, and recent transformer‑ and foundation‑based multimodal approaches are omitted from comparison.
- 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.
(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?
This is a technically solid and well‑motivated study that advances multi‑task glioma classification through thoughtful module design and strong validation. The paper is sufficiently novel, clearly presented, and experimentally convincing.
- 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.
I maintained my acceptance recommendation because the paper’s strengths outweigh its limitations, and the concerns raised in the initial review were sufficiently clarified in the rebuttal or left as reasonable directions for future work.
Review #2
- Please describe the contribution of the paper
The paper proposes DEC-Net, a unified multi-task framework for jointly predicting glioma grading and molecular subtypes from multimodal MRI, aiming to address inter-task interference while leveraging shared representations. It introduces a progressive mixture-of-experts module to balance shared and task-specific feature learning, enabling more effective multi-task optimization. In addition, the method incorporates a dependency-aware loss to explicitly model the biological correlations among tasks, encouraging more consistent and clinically plausible predictions.
- 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.Strong Empirical Performance The proposed method demonstrates consistent performance improvements over existing approaches on both internal and external datasets. Notably, the model maintains its advantage even under domain shift, suggesting a certain level of robustness and generalization capability.
2.Public Availability of Code The authors provide publicly available code, which enhances the reproducibility of the work and facilitates future research building upon this method.
3.Comprehensive Ablation Study The paper includes a thorough ablation study evaluating the contribution of each proposed component. This analysis helps demonstrate that the individual modules are complementary and collectively contribute to the overall performance improvement.
- 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.Insufficient Analysis of Expert Routing Dynamics While the model successfully separates shared and task-specific experts, an in-depth analysis of their individual contributions is lacking. Providing a more granular investigation into how these experts are utilized across different tasks, along with an interpretability analysis of the routing mechanism, would significantly strengthen the justification for this architectural design.
2.Architectural Complexity and Presentation Clarity The proposed framework integrates multiple intricate modules, resulting in a relatively complex architecture that can be challenging to follow. While space limitations are acknowledged, a more streamlined explanation or supplementary visualizations detailing the data flow through these key components would greatly improve the paper’s readability and reproducibility.
3.Limited Scope of Baseline Comparisons The current experimental comparisons are primarily restricted to multi-task learning (MTL) methods. To comprehensively evaluate the proposed approach, the authors should include strong single-task baselines or more recent state-of-the-art architectures. This would rigorously demonstrate whether the performance gains stem from the multi-task synergy itself rather than merely the overall model capacity.
- 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 tackles a crucial clinical challenge with a well-designed multi-task framework (DEC-Net). The major factors driving my positive score are the model’s strong, consistent empirical performance across both internal and external datasets (demonstrating impressive robustness to domain shift), the thorough ablation study, and the provision of public code for reproducibility.
However, I am leaning toward a Weak Accept rather than a Strong Accept due to a few analytical and structural limitations: 1.Lack of deep analysis: The paper needs a more granular, interpretability-focused analysis of how the PMoE routing dynamically allocates shared and task-specific experts. 2.Presentation clarity: The architecture is quite complex, and the manuscript would benefit from a more streamlined explanation or clearer visual flow of the components. 3.Missing single-task baselines: Comparing only against other multi-task methods leaves it unclear whether the performance gains stem from true multi-task synergy or simply an increase in overall model capacity.
Overall, despite these gaps in baseline comparisons and interpretability, the robust experimental results, public availability, and significant clinical potential make this paper a valuable contribution to the field.
- Reviewer confidence
Somewhat confident (2)
- [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 appreciate the authors’ clear and concise rebuttal. My previous concerns regarding the limited baseline comparisons and the lack of expert routing dynamics have been sufficiently addressed with the newly provided single-task performance results and the specific shared/task-specific routing ratios. I am satisfied with these clarifications and happily confirm my final decision to Accept this paper.
Review #3
- Please describe the contribution of the paper
The authors propose DEC-Net, a multi-task deep learning framework for joint glioma grading and molecular subtyping from multimodal preoperative MRI (T1, T1ce, T2, FLAIR). The framework comprises three components: (i) a Distillation ResModule (DRModule) in which an EMA-updated teacher branch, gated by a spatial discrepancy map obtained from an unsupervised anomaly-detection pre-training stage, distils multi-scale features into a student ResNet; (ii) a Progressive Mixture-of-Experts (PMoE) module with shared and task-specific experts and per-task softmax gating under dense routing, stacked across L layers; and (iii) a Dependency Loss (D-Loss) derived from an empirical class-conditional co-occurrence matrix P computed on the training set, intended to encode biological correlations among the three tasks. Training uses focal loss combined with D-Loss and the distillation loss, with fixed balancing coefficients. Evaluation is performed via 5-fold cross-validation on a combined UCSF-PDGM + EGD internal cohort and held-out external testing on TCGA-GBM/LGG, benchmarked against nine multi-task glioma methods using AUC and accuracy. Anonymised source code is provided.
- 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.
-
Clinically meaningful joint formulation. Non-invasive preoperative prediction of glioma grade together with IDH mutation and 1p/19q co-deletion is of genuine clinical value, and jointly modelling the three tasks respects the biological coupling among the canonical diagnostic categories.
-
Substantial multi-centre cohort with a genuinely held-out external test set. Training and cross-validation span two institutions (UCSF-PDGM and EGD, total n=1,214), and TCGA-GBM/LGG (n=167) is used exclusively as an external cohort. Many prior works report single-centre results only, and it enables a meaningful assessment of domain shift.
-
External-test generalisation. On TCGA, DEC-Net leads the best competing baseline. Several baselines (e.g., MTAFNet, MFEFNet, MLNet) experience substantial AUC drops under domain shift, whereas DEC-Net degrades more gracefully. The relative ranking of methods also changes between internal and external evaluation, which is consistent with a genuine test of generalisation.
-
Complete module ablation. Table 3 enumerates all eight on/off combinations of DRModule, PMoE, and D-Loss rather than the more common single-leave-one-out design.
-
Reproducibility. Preprocessing is specified (SRI24 co-registration, 1 mm isotropic resampling, skull-stripping, z-score normalisation), optimiser (AdamW), learning rate (1e-4), batch size (8), epoch count (100), and loss weights (alpha=0.3, beta=0.1) are all reported. An anonymised code repository is linked.
-
- 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.No statistical significance testing. Table 2 reports mean and standard deviation across cross-validation folds, but no paired statistical test (DeLong for AUC, McNemar for accuracy, or paired Wilcoxon on fold-level scores) is provided. Several internal-test “wins” fall within one standard deviation of the runner-up. Without paired testing, the claim of “consistently outperforms SOTA” on the internal cohort is not empirically supported.
2.Handling of a large fraction of Unknown labels is not stated, with a plausible confound in the D-Loss. Table 1 shows that IDH status is Unknown for 42% of EGD and 11% of TCGA, and 1p/19q status is Unknown for 68% of EGD and 18% of UCSF-PDGM. The paper does not state how these cases are handled during training and evaluation. This matters on two fronts. First, the effective per-task labelled sample sizes are materially smaller than the headline n=1,214; for 1p/19q in particular the labelled pool may be well below 700.Second, if P is estimated only from fully-labelled patients, it is computed on a biased subsample.
3.Unspecified fold stratification. UCSF-PDGM and EGD are acquired on different scanners and protocols. A random patient-level split mixes sites across train and validation folds, which allows the model to exploit site-specific cues (residual intensity distributions, etc.) and can inflate internal-test metrics. The paper does not state whether folds are patient-level only, site-stratified, or class-stratified, nor does it report per-fold site composition.
4.Unverified fairness of the nine SOTA baselines is unverified. The paper does not state whether the baselines were re-implemented in-house, whether they were trained under the same preprocessing pipeline, input size (128×128×128), modality set, and augmentation, whether they received the same hyperparameter-tuning budget, or how they handle Unknown labels. Several baselines (MTTU-Net, MDLNet, CMTLNet, MTAFNet, MLNet) are multi-task glioma methods whose original publications use different cohorts and label subsets. Simple re-implementation could systematically disadvantage them.
5.Methodological novelty is incremental. The EMA teacher-student mechanism is the Mean Teacher scheme of Tarvainen and Valpola (NeurIPS 2017) not the cited reference [27] which is a 2025 application paper. Progressive Layered Extraction (PLE) is not cited or contrasted. The D-Loss is a label-correlation regulariser from the multi-label classification literature. The paper’s contribution is therefore the combination applied to joint glioma profiling, which is a legitimate application-style contribution, but the current framing presents each component as novel without acknowledging the closest prior art.
6.Unexplained interactions in the ablation. In Table 3, the “PMoE + D-Loss” configuration underperforms “D-Loss only” on Grade AUC (91.68 vs 92.63) and 1p/19q AUC (82.37 vs 84.77), while the full three-module model then jumps to 95.11 and 85.45 respectively. This is a super-additive effect that is not discussed.
7.No calibration or operating-point analysis. For a method positioned to “support precise clinical decision-making”, calibration and a sensitivity-at-fixed-specificity table would strengthen the clinical framing.
- 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
- Table 2 header: “Interal Test” should read “Internal Test”.
- Preprocessing omits N4 bias-field correction, which is standard for harmonising multi-site brain MRI. Worth commenting on whether this was tested and found not to help, or simply omitted.
- Reference [27] (EMA) cites a 2025 self-distillation paper rather than Tarvainen, Antti, and Harri Valpola. “Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results.” Advances in neural information processing systems 30 (2017)., which is the canonical reference. Consider adding the original source.
- The claim that DRModule “preserves deployment efficiency” would be strengthened by a FLOPs and parameter-count comparison against the plain student backbone, since the teacher is used only at training time.
Rebuttal questions (to focus the authors’ response): Q1.How are Unknown IDH (EGD 42%, TCGA 11%) and Unknown 1p/19q (EGD 68%, UCSF 18%) cases treated during training, during evaluation, and importantly during estimation of the conditional matrix P in Eq. 8? What are the effective per-task labelled sample sizes?
Q2.CV stratification: are the five folds patient-stratified only, or also site-stratified across UCSF-PDGM and EGD? Please provide per-fold site composition.
- 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 addresses a clinically meaningful problem (non-invasive joint prediction of glioma grade, IDH mutation, and 1p/19q co-deletion) on a larger-than-typical multi-centre cohort with a genuinely held-out external test set. The external-test margins over strong multi-task baselines are the most convincing part of the empirical case, and the ablation design is above the typical standard.
The paper falls short of the acceptance threshold on four grounds.
First, no statistical significance testing is provided despite several internal-test wins that fall within one standard deviation of the runner-up. Paired DeLong and McNemar tests on existing fold-level scores would substantially strengthen the case and are answerable in the rebuttal.
Second, handling of the large Unknown-label fractions is not stated. This affects both the effective training sample size and more seriously the estimation of the conditional probability matrix P in Eq. 8, which is the core of the proposed D-Loss. If Unknowns are dropped from P estimation, the prior is computed on a biased subsample and this bias propagates into every gradient update.
Third, the 5-fold CV stratification on the combined UCSF-PDGM + EGD cohort is not specified. Multi-site MRI with random patient-level splits carries non-negligible site-leakage risk, and without site-stratified or leave-one-site-out evidence the internal-test margins cannot be cleanly interpreted as generalisation rather than site exploitation.
Fourth, methodological novelty is incremental: the EMA teacher-student scheme is Mean Teacher, PMoE closely tracks Progressive Layered Extraction (PLE which was not cited), and D-Loss is a standard label-correlation regulariser. The combination applied to joint glioma profiling is a legitimate application-style contribution, but the current framing presents each component as novel.
- 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 two most structurally serious concerns raised in my initial review have been addressed to a satisfactory degree. The authors confirmed that samples with Unknown labels were excluded from training, evaluation, and P-matrix estimation, and provided the effective per-task labelled sample sizes (Grade n=1,166; IDH n=910; 1p/19q n=632). The confirmation that 5-fold cross-validation is site-stratified across UCSF-PDGM and EGD substantially mitigates the site-leakage concern I had raised. Combined with the genuinely held-out TCGA external test set, the validation design is now sufficiently credible.
The authors also report that DeLong and McNemar tests were conducted, with p < 0.05 for several AUC comparisons and all ACC comparisons against baselines. While these figures are not present in the submitted manuscript, the directional claim is plausible given the fold-level standard deviations in Table 2. Concerns that remain unresolved: the super-additive ablation interaction between PMoE and D-Loss is not discussed; the EMA/Mean Teacher and PLE citations were promised but are not in the submitted manuscript; and calibration figures on the external cohort are absent. These are appropriate targets for the camera-ready revision.
Author Feedback
We thank the reviewers and meta-reviewer (MR) for their constructive comments. We address their concerns below. 1.Model Complexity and Performance(MR, R2Q2, R3Q1): The pre-trained network extracts tumor-relevant regions, the distillation framework enhances feature representation, the expert network mitigates inter-task interference, and the biologically guided D-Loss optimizes multi-task prediction. Together, these components improve glioma grading and molecular subtyping. Statistical tests on internal data show significant improvements: for AUC, significant differences are observed compared with several baseline methods (DeLong test, p < 0.05); for ACC, improvements are significant compared with all baseline methods (McNemar test, p < 0.05). 2.Rationale for Not Adopting Existing Tumor Segmentation Methods(MR): Gliomas affect the whole brain environment. Focusing only on tumor segmentation masks is insufficient to capture the complete affected regions. Our approach captures comprehensive tumor-relevant areas without relying solely on existing segmentation methods. 3.Sensitivity (SEN) and Specificity (SPE) for Our Task (MR,R3Q7): Internal test SEN/SPE: Grade 88.14 ± 1.83 / 91.55 ± 2.47; IDH 84.99 ± 5.57 / 85.47 ± 2.79; 1p/19q 79.48 ± 5.25 / 79.89 ± 4.28.Due to space limits, these metrics were omitted from main tables but demonstrate superior performance over other methods. 4.Implementation and Inference Details (R1Q1): The center cropping effectively captures the core brain area while removing boundaries. During inference, we perform whole-volume prediction rather than patch-based aggregation. 5.Baseline Comparisons and Limitations (R1Q2, R2Q3): Our core contribution lies in multi-task learning , and our experiments mainly compare with recent SOTA multi-task methods. In our experiments, single-task results were evaluated. For ACC, multi-task synergy outperforms single-task baselines (Grade ACC 85–87% , IDH ACC 80–82%, and 1p/19q ACC 72–74%). We will add a brief discussion of limitations in the revised version. 6.Expert Routing Analysis (R2Q1): Full visualization of expert contributions is omitted due to page limits. We instead report the ratio of shared/task-specific expert selection: Grade (86/14), IDH (74/26), and 1p/19q (68/32). These results indicate that each task leverages different experts while showing task-related patterns, supporting the effectiveness of our routing mechanism. 7.Handling of Missing Labels (R3Q2): Samples with missing labels were excluded from model training, P-matrix estimation, and evaluation. We did not adopt pseudo-labeling, as unreliable pseudo-labels may introduce noisy prior knowledge and lead to model hallucination or error accumulation. As shown in Table 1, the valid annotated sample sizes are 1,166 for Grade, 910 for IDH, and 632 for 1p/19q. These counts are used for P-matrix estimation. 8.Cross-Validation Stratification (R3Q3): Our five-fold cross-validation is site-stratified. Each fold contains one-fifth of samples from both UCSF-PDGM and EGD, ensuring that data from different scanners and protocols are represented in every fold. Additionally, (1) bias-field correction was applied during preprocessing, (2) there are no strong site-specific correlations that could inflate metrics, and (3) external validation confirms the model’s robustness. 9.Fairness of Baseline Comparisons (R3Q4): To ensure a fair comparison, all nine SOTA baseline methods were evaluated under identical experimental settings, including the same preprocessing pipeline, fixed input size, consistent modality combination. 10.Citation and Methodology (R3Q5): We will add proper citations for EMA teacher-student and PLE. In our work, these components are adapted for glioma tasks, considering task correlations. Specifically, we introduce a shared-and-specific structure and a D-Loss guided by biological priors, which together constitute a novel application-oriented contribution for multimodal glioma analysis.
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.
The paper proposes a multi-sequence MRI-based framework for glioma grading and molecular subtyping. However, the overall architecture appears overly complex, while only marginal improvements are observed compared with existing methods. To capture tumor-relevant regions, it is unclear why existing well-established tumor segmentation methods are not adopted, as many of them can already achieve highly accurate tumor delineation (e.g., Dice score > 0.9). In addition, class imbalance (e.g., IDH status) is a well-known challenge in glioma. Although focal loss is employed, it would be helpful to further report accuracy and recall for IDH mutation and 1p/19q co-deletion, to better assess model performance.
- 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 concerns regarding the paper have been addressed, and all reviewers agree to accept the paper.
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.
The paper proposes DEC-Net, a multi-task framework for joint glioma grading, IDH prediction, and 1p/19q prediction from multimodal MRI. All reviewers accepted the paper after rebuttal, which appears to have addressed the main concerns sufficiently.
However, the authors should more carefully address class imbalance by including additional metrics, such as per-class performance, sen/spe, in the camera-ready version. They should also include statistical analyses for the comparisons and more clearly position their technical contributions relative to the existing literature. In addition, rows 4, 5, 6 of the ablation study show that the roles of DRModule and PMoE are not fully evident, and this should be better discussed.
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.
This paper proposes DEC-Net, a multi-task framework for glioma grading and molecular subtyping from multimodal MRI. Reviewers appreciated the clinical relevance, strong performance on both internal and external cohorts, and the comprehensive ablation studies. I agree with the reviewers that the method is technically sound and well validated. The rebuttal satisfactorily addressed concerns regarding unknown-label handling, cross-validation design, statistical significance testing, and baseline fairness. While some components build upon existing methods, the overall framework demonstrates clear practical value and strong generalization ability. Therefore, I recommend acceptance.
