List of Papers Browse by Subject Areas Author List
Abstract
Deep learning-based survival analysis with the Cox proportional hazards loss is challenged in 3D medical imaging due to GPU memory constraints, which enforce small mini-batches and thus yield an insufficient risk set and noisy optimization. To address this issue, we propose a Momentum-Contrastive Survival Framework (MCSF) that maintains a momentum-updated virtual cohort bank, decoupling the effective risk set size from the mini-batch size and enabling a larger global risk pool across iterations. Beyond enlarging the risk pool, we introduce a time-aware ranking objective that explicitly leverages event-time information: each event sample serves as an anchor and is compared against other event samples with temporal proximity-based weighting, providing a more stable and informative ranking signal than batch-limited Cox optimization. MCSF can be seamlessly integrated with various 3D backbones for end-to-end learning. Experiments on Breast and Head & Neck cancer cohorts demonstrate consistent improvements, achieving C-indices of 0.716 and 0.654, respectively, outperforming strong baseline backbones and highlighting the robustness of MCSF for 3D PET-CT survival prediction. Code is available at https://github.com/XinglongLiang08/MCSF.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3495_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/XinglongLiang08/MCSF
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiaXin_Beyond_MICCAI2026,
author = { Liang, Xinglong AND Zhang, Tianyu AND Cai, Lishan AND Huang, Jiaju AND Lu, Chunyao AND Han, Luyi AND Duan, Yaofei AND Wang, Xin AND Gao, Yuan AND He, Muzhen AND Xu, Jun AND Sun, Yue AND Tan, Tao AND Mann, Ritse},
title = { { Beyond the Batch: Momentum-Updated Virtual Cohorts for 3D PET-CT Prognosis } },
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
1.This paper identifies an important optimization bottleneck in 3D PET-CT survival learning, where GPU memory constraints force very small batch sizes and limit the effectiveness of Cox-style risk modeling. To address this, the paper proposes a Virtual Cohort Bank (VCB) to approximate a larger global risk set during training. 2.The paper further introduces a time-aware weighted ranking objective (TAPRL), which aims to improve global risk ranking by down-weighting ambiguous patient pairs with similar survival times. 3.The proposed framework is validated on two clinically relevant survival prediction tasks, namely breast cancer and head and neck cancer, where it shows promising performance improvements and suggests a practical direction for more stable survival modeling from 3D medical images.
- 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.
Addresses a practically important bottleneck: The paper tackles a real limitation of 3D medical survival modeling, namely the mismatch between memory-intensive full-field-of-view 3D imaging and the large effective risk sets preferred by survival objectives. The proposed framework offers a reasonable way to alleviate this issue without relying on aggressive cropping or downsampling that may discard clinically relevant context.
Reasonably general framework: The proposed Momentum Contrast Survival Framework (MCSF) appears to be reasonably model-agnostic and can be integrated with multiple common 3D backbones for end-to-end learning. The ablation experiments suggest that the gains are not restricted to a single encoder design.
Thoughtful treatment of historical features: The use of an evolving reference encoder is a sensible design choice for mitigating the instability that may arise when reusing historical representations. In addition, the time-aware weighting mechanism is well motivated and helps reduce the influence of ambiguous pairwise comparisons.
Consistent empirical improvements: The reported results on two distinct cancer cohorts are encouraging and suggest that the method can improve survival prediction performance over the included baselines. The Kaplan-Meier stratification results further support the practical utility of the approach.
- 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 model’s input pre-processing stage relies on a binarised lesion segmentation mask as a spatial prior, applying affine transformations and reweighting to the PET and CT channels. As the datasets used in the experiments all contain pre-existing segmentation masks, this implies that, for practical clinical deployment, the framework must be paired with a high-precision automatic segmentation model or rely on manual annotation. 2.This paper suffers from shortcomings in terms of experimental rigour and persuasiveness; it lacks quantitative data such as the number of parameters, FLOPs and GPU memory usage. This not only contradicts its central claim of ‘resolving computational bottlenecks’, but also fails to provide quantitative evidence to account for the additional computational overhead incurred by training two networks in parallel. 3.The empirical comparison remains somewhat limited, as the main baselines are predominantly generic 3D CNN backbones and an NLL variant, while direct comparisons against stronger survival-specific models are missing. This makes it difficult to isolate whether the gains come primarily from the proposed survival formulation or from favorable backbone/task settings.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
Overall, I find this paper technically interesting and potentially useful for 3D PET-CT survival prediction. The paper identifies a meaningful optimization bottleneck in small-batch Cox-based survival learning and proposes a reasonable solution by enlarging the effective risk pool through a virtual cohort bank and a momentum-updated reference encoder. The introduction of a time-aware ranking objective is also well motivated, and the reported results on two cancer cohorts suggest that the approach can provide consistent performance improvements.
The main strengths of the paper are: (1) it addresses a practically relevant limitation of 3D survival modeling under GPU memory constraints; (2) the proposed framework is reasonably general and can be integrated with different backbone architectures; and (3) the empirical results indicate consistent gains across datasets and backbones.
At the same time, I do have several reservations. The method depends on lesion masks as an explicit prior, which may limit real-world applicability unless high-quality segmentation is available. In addition, the paper does not provide sufficient quantitative evidence regarding computational overhead, despite positioning the method as a response to memory-constrained training. Finally, the comparison with stronger survival-specific baselines is limited, which weakens the overall empirical claim.
I therefore view this paper as marginally above the acceptance threshold. The core idea is meaningful and the results are promising, but the paper would be stronger with more rigorous experimental validation and clearer discussion of deployment-related limitations.
- 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 rebuttal, I lean toward acceptance.
The authors have addressed several of the main concerns in a helpful way. In particular, they now provide quantitative clarification of the computational overhead, clarify that the same mask-guided input modulation was used across baselines, explain the queue mechanism more clearly, and give additional information about the private cohort. They also make it clear that the Kaplan–Meier analysis should be treated as supportive qualitative evidence rather than a strict basis for superiority claims. These points make the paper clearer and address some of my earlier concerns about technical presentation and evaluation fairness.
Some issues are still not fully resolved. In particular, the rebuttal does not completely disentangle the contribution of TAPRL from that of the enlarged virtual cohort, and the comparison with other Cox approximation strategies remains limited. Even so, I find the problem important and practically relevant, and I think the rebuttal addresses the main concerns well enough for a borderline positive decision.
Review #2
- Please describe the contribution of the paper
The paper addresses survival prediction from 3D PET-CT in the small-batch regime. The main idea is to enlarge the effective comparison pool for risk ranking by combining the current mini-batch with a momentum-updated virtual cohort bank, and to further weight pairwise comparisons by survival-time separation. Experiments on breast cancer and head-and-neck cancer cohorts show improved C-index and risk stratification over several CNN baselines.
- 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 tackles a relevant practical issue: in 3D PET-CT survival modeling, small mini-batches can make Cox-style risk-set estimation unstable. The proposed virtual cohort bank is intuitive and fits the application setting well. The time-aware weighting is a reasonable way to reduce ambiguous comparisons. Results on two cohorts are promising, and the ablation suggests both VCB and TAPRL contribute.
- 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 novelty is somewhat limited from the current presentation. The method appears to combine momentum encoder, memory bank, and pairwise ranking ideas in a survival setting, but the distinction from related momentum-contrastive / ranking-based work is not yet very clear. The role of the lesion mask prior is unclear. The method uses mask-guided input modulation, but I could not tell whether all baselines received the same mask information, or whether this component was ablated separately. This makes attribution of the gains harder. The need for full-volume training is asserted rather than fully justified. It would help if the paper clarified why ROI/crop-based alternatives are not appropriate in this setting. Reproducibility is limited by missing implementation details of the queue: what exactly is stored, whether queued scores/features are refreshed, and whether both event and censored samples are enqueued. The computational tradeoff is not discussed. Since the method adds a reference encoder and a virtual bank, some comment on runtime/memory overhead would help.
- 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 find the problem important and the proposed idea plausible, especially for very small-batch 3D PET-CT training. However, I currently lean slightly negative because the paper does not yet fully separate the contribution of the proposed framework from the lesion-mask prior, and some important methodological details are underspecified. In addition, the novelty relative to existing memory-bank / ranking-style ideas is not yet fully clear from the paper. Overall, I see this as a borderline submission with a reasonable practical idea, but with some remaining concerns about novelty, fairness, and reproducibility.
- 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 rebuttal, I think the authors have addressed most of my main concerns. They clarified the computational cost, noting that the reference encoder is EMA-updated without backpropagation and that inference uses only the main encoder. They also clarified that all baselines used the same mask-guided input, which makes the comparisons fairer.
The rebuttal also provides useful details about the virtual cohort bank, including what is stored, how the queue is updated, and that both event and censored samples are included. The explanation for using full-volume PET-CT rather than ROI cropping is reasonable.
I still think the novelty is moderate, since the method combines momentum encoders, memory banks, and ranking losses. Also, the contribution of the time-aware weighting would be clearer with a flat-weighted VCB ablation. However, the paper addresses an important small-batch problem in 3D PET-CT survival modeling and shows consistent empirical improvements. Overall, I support acceptance.
Review #3
- Please describe the contribution of the paper
This paper proposes the Momentum-Contrastive Survival Framework (MCSF), which addresses the small-batch bottleneck in Cox proportional hazards-based survival learning for 3D PET-CT data. The core idea is adapted from MoCo: a momentum-updated Evolving Reference Encoder generates stable feature representations that are stored in a Virtual Cohort Bank (VCB), decoupling the effective risk set size from GPU-constrained mini-batch size.
- 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 mismatch between the global nature of Cox partial likelihood and the severely constrained batch sizes in 3D medical imaging is a genuine and under-addressed issue. 2.able 2 evaluates VCB and TAPRL separately across three backbones on both datasets, showing consistent additive gains. Table 3 examines VCB size sensitivity, including the non-monotonic degradation at K=128, which is an honest and informative result. 3.Using lesion masks as a learnable spatial prior for PET/CT weighting is a sensible way to focus the backbone on relevant regions without hard masking.
- 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 larger and more impactful dataset (1,210 patients) is from a private collection. No dataset characteristics (cancer subtype, staging, treatment, event rate, follow-up duration) are reported. This makes reproducibility and generalizability of the main results unverifiable. 2.TAPRL combines (a) enlarging the risk pool via VCB and (b) time-aware weighting. Table 2 ablates VCB vs VCB+TAPRL but does not separately evaluate the time-aware weighting against a flat-weighted version of the enlarged risk set. It is therefore unclear whether the temporal weighting provides benefit beyond simply having a larger risk pool. 3.The paper compares only against backbone architectures trained with Cox or NLL loss on the in-batch risk set. There is no comparison against other large-risk-set Cox approximations. This significantly limits the positioning of MCSF relative to prior work.
- The momentum encoder is justified by analogy to MoCo, but the degree of feature drift in the VCB and its impact on risk score consistency across iterations is never analyzed. 5.Panels d and e (ResNet34 and SEResNeXt50) show p > 0.05 for log-rank tests, which is presented without comment. The comparison between methods via KM significance is not a reliable metric for model comparison, and framing it as evidence of superiority is misleading.
- 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.
(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?
MCSF addresses a genuine and well-motivated problem with a clean adaptation of momentum-contrastive learning to Cox survival optimization. The repeated experiments with variance reporting, consistent ablation results across multiple backbones, and evaluation on two cancer types support the core claims. The main weaknesses are the private dataset with inadequate characterization, the absence of comparison against existing Cox approximation methods, and insufficient disentanglement of VCB vs. TAPRL contributions. These are addressable in rebuttal.
- 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 have addressed my comments, so i am raiising my score.
Author Feedback
We thank all reviewers(R) and have responded with the necessary information.
Q1: Computational cost.(R1R2) A1: Taking ResNet-18 as an baseline, MCSF increased parameters from 8.3M to 16.6M, forward FLOPs from 285G to 570G, and GPU memory from 6036M to 7686M. The reference encoder is EMA-updated without gradient backpropagation, limiting additional training overhead. Inference uses only the main encoder, cost equals baseline.
Q2: Mask and ROI.(AC,R1R2) A2: All compared models used the same mask-guided input modulation, ensuring fair comparison. Full-volume training was adopted because PET-CT lesions are spatially sparse and may include metastatic lymph nodes and even occult metastasis; ROI cropping may miss clinically relevant disease. Internal masks were generated by our in-house segmentation model and clinically reviewed. Missing H&N masks were inferred using the open-source weights of the 2025 H&N challenge winner. We will release our segmentation weights.
Q3: Baseline Selection and Survival Models.(R1R4) A3: ResNet with Cox or NLL losses remains a mainstream paradigm for 3D medical prognosis. We focused on the risk-set bottleneck within this standard framework to isolate MCSF’s contribution. Consistent gains over these established baselines support our framework’s efficacy.
Q4: Novelty.(AC,R2)： A4: To the best of our knowledge, MCSF is the first framework to integrate momentum-based memory mechanisms with survival theory to address the small-batch bottleneck in 3D imaging. Unlike conventional momentum-contrastive methods that store features for instance discrimination, MCSF stores risk scores, event times, and labels to approximate a global virtual cohort. This enables each event sample to be ranked beyond the local mini-batch. TAPRL further leverages temporal separation to reduce noisy pairs, making the framework specifically tailored to survival analysis rather than a simple combination of existing methods.
Q5: Details for queue.(AC,R2) A5: The queue stores risk scores, event times, and labels. After each batch, the latest patients are enqueued and the oldest samples are removed. Both event and censored patients are enqueued. Existing queued scores are not re-computed; their consistency is maintained by the EMA encoder.
Q6: Private Dataset.(AC,R4) A6: The private cohort included I–III stage patients. The maximum follow-up time was 13 years; the event rate was 9.8%. The data includes all subtypes, and treatments include surgery and neoadjuvant therapy.
Q7: Ablation for TAPRL.(R4) A7: TAPRL was designed as a refinement of VCB rather than an independent module: VCB restores sufficient cross-patient comparisons under small-batch 3D training, while TAPRL reduces noise by down-weighting temporally ambiguous pairs. Currently, we justify this design theoretically, and a flat-weighted VCB ranking ablation will be included in future work.
Q8: Other large-risk-set Cox approximations.(AC,R4) A8: Existing methods mainly target low-dimensional features, where historical samples can be revisited. This is impractical for 3D PET-CT due to memory constraints. MCSF instead stores EMA-derived scores, enlarging the risk set without backpropagating historical 3D volumes.
Q9: Feature drift.(R4) A9: The EMA encoder aims to smooth parameter updates and maintain more consistent stored risk scores. Table3 indirectly suggests the potential impact of feature drift: performance drops when increasing VCB from K=64 to 128, likely due to older representations. We will clarify this rationale in the revision.
Q10: KM interpretation.(R4) A10: We agree that KM analysis is related to risk discrimination but does not strictly correspond to C-index. In the revision, we will explicitly note that panels d and e show non-significant log-rank results, indicating limited separation for these two baseline models, and present KM curves only as supportive qualitative evidence of risk stratification.
These essential details will be clarified in the revision.
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.
Although two reviewers gave an accept rating, one reviewer leaned negative due to insufficient separation of the framework’s contribution from the lesion-mask prior, underspecified methodological details, and unclear novelty relative to existing memory-bank or ranking-based approaches. They also mentioned weaknesses, including an inadequately characterized private dataset and a lack of comparison with existing Cox approximation methods. The authors should clearly defend their paper against these issues.
- 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 authors have adequately addressed all of the reviewers’ concerns. Accordingly, I recommend acceptance of 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 authors have clarified several of the reviewers’ concerns, notably the computational overhead and experimental protocol. Despite the remaining shortcomings, there is a consensus for 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.
The reviewers unanimously provided positive recommendations, and I agree with their decision.