Abstract

Despite the success of deep learning models to detect clinically significant prostate cancer (csPCa) using multiparametric MRI (mpMRI), it faces two levels of data imbalance: (1) voxel-level class imbalance within each patient case, where lesion voxels occupy a small fraction of the image volume, and (2) case-level difficulty imbalance, where cases exhibit heterogeneity in detection difficulty. While existing methods primarily address voxel-level imbalance via loss reweighting, they often overlook the uneven distribution of case difficulty, limiting their performance. To bridge this gap, we propose Reverse Symmetric Self-Paced Learning (RSSPL), a model-agnostic case-level learning strategy. Specifically, RSSPL builds upon curriculum learning but prioritises patient cases in a hard-to-easy manner to better emphasise challenging cases earlier during training, and can be used alongside voxel-level reweighting. To facilitate this, an anatomy-aware difficulty estimate is proposed to refine case scheduling by filtering the noise outside the prostate gland. Experiments on csPCa detection across different network backbones and both standard and imbalance-aware loss functions demonstrate that RSSPL leads to reduced cross-validation variance and improved detection performance. On the independent Prostate158 test set, RSSPL achieves a statistically significant improvement at the patient level, with an absolute AUROC increase of 3.05% (p < 0.05). Consistent improvement trends in a brain lesion-detection task further provide preliminary evidence of RSSPL’s cross-organ transferability.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2755_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{ZhaDan_AnatomyAware_MICCAI2026,
        author = { Zhang, Dan AND Zhao, Di AND Sun, Yu AND Wang, Alan AND Aguergaray, Claude AND Reynolds, Hayley M.},
        title = { { Anatomy-Aware Reverse Symmetric Self-paced Learning for Prostate Cancer Detection } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes Reverse Symmetric Self-Paced Learning (RSSPL), a case-level training strategy for clinically significant prostate cancer detection on mpMRI. The method aims to address case-level difficulty imbalance rather than voxel level by reversing the conventional easy-to-hard curriculum and instead prioritizing harder cases early in training. In addition, the paper introduces an anatomy-aware difficulty estimation scheme, where gradient-based difficulty is computed within the prostate gland region using anatomical masks, in order to reduce background noise in difficulty ranking. The method is evaluated across multiple baselines and loss functions, with additional validation on an external prostate dataset and a secondary brain lesion task.

  • 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 addresses a clinically meaningful and underexplored problem in prostate mpMRI detection, case-level difficulty imbalance in addition to the more commonly studied voxel-level class imbalance. The proposed RSSPL framework is simple to understand and it can be integrated with different loss functions and model architectures. I also find the anatomy-aware difficulty estimation reasonable and clinically motivated, as it restricts the gradient-based difficulty signal to the prostate gland rather than the whole image, which helps reduce irrelevant background noise.

    The experimental evaluation includes multiple backbones and loss settings, an independent test set on Prostate158, and statistical validation with paired bootstrap analysis. In addition, the qualitative example in Fig. 2 is particularly compelling: cross-sequence misalignment is a realistic and important challenge in prostate MRI (misalignment between different MR modalities).

    Overall, this is a well-motivated and practically useful training strategy with consistent empirical benefits.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    While the paper is well motivated, it has several limitations. The improvements are relatively modest in several configurations, and the statistically significant gain is mainly observed for patient-level AUROC rather than lesion-level. The generalisation claim is not yet fully convincing, since the cross-organ validation is conducted on a simplified brain lesion setting with a different difficulty signal and saturated AUROC. Finally, the paper does not sufficiently discuss the additional computational overhead introduced by per-epoch gradient-based difficulty estimation and global case ranking.

  • 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?

    I chose weak accept because the paper presents a meaningful and reasonably novel idea for handling case-level difficulty imbalance, and the method is simple, model-agnostic, and supported by fairly solid experiments. However, I do not score it higher because the improvements are modest in several settings, the strongest significant gain is mainly on patient-level AUROC, and the generalisation evidence is still limited.

  • 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.

    N/A



Review #2

  • Please describe the contribution of the paper

    The primary contribution of this work is the introduction of Reverse Symmetric Self-Paced Learning (RSSPL), a novel curriculum training strategy designed to address polarized case-level difficulty in medical imaging datasets. The authors propose an anatomy-aware difficulty estimation, utilizing a clinically informed, gradient-based measure that strategically restricts focus to the relevant prostate anatomy. Furthermore, the paper establishes the robust generalizability of the proposed approach by demonstrating its application across distinct medical domains, yielding consistent performance improvements in both prostate cancer and brain lesion detection tasks.

  • 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 introduction of Reverse Symmetric Self-Paced Learning (RSSPL) offers an original, plug-and-play curriculum strategy. Because the method is entirely model-agnostic, it can be integrated with standard architectures, such as U-Net and MNet, as well as conventional loss functions. This simplicity is a significant strength, as it allows for broad adoption without requiring complex, custom architectural modifications.
    • The decision to restrict the gradient-based difficulty measure specifically to the prostate gland is methodologically clever and justified from a clinical perspective. By isolating the relevant anatomical signal, the method avoids confounding background noise. This design choice is supported and empirically validated through the authors’ ablation studies.
    • The paper identifies and addresses a frequently overlooked bottleneck in medical image segmentation: case-level difficulty imbalance. By shifting the focus away from the more commonly studied voxel-level imbalance, the authors tackle a problem that directly impacts the training dynamics of real-world clinical datasets.
    • The experimental setup is robust. The authors utilize a large-scale, multi-center dataset (PI-CAI) alongside an independent external testing cohort. The inclusion of statistical significance testing, combined with a separate study demonstrating the method’s generalizability across different tasks, strongly validates the performance claims.
    • The results show a measurable reduction in cross-validation standard deviation. This proven improvement in model stability and reproducibility is a crucial prerequisite for the actual clinical deployment and reliability of medical AI systems.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • While intuitively motivated for polarized data distributions, the paper fails to rigorously establish why a “hard-to-easy” curriculum fundamentally outperforms the traditional “easy-to-hard” approach (which has been the standard since Bengio et al., 2009). The claims would be significantly strengthened by deeper mathematical grounding or, at a minimum, controlled synthetic experiments demonstrating the failure modes of easy-to-hard pacing on polarized data.
    • The anatomy-aware difficulty estimation inherently requires high-quality prostate segmentations. In real-world clinical pipelines, perfect masks are rarely guaranteed, and automated segmentation often introduces artifacts. The authors completely omit a sensitivity analysis evaluating how segmentation noise or inaccurate masks degrade the difficulty estimation and subsequent model performance.
    • Difficulty scores are computed only once per epoch. Because model gradients and parameter states evolve continuously during training, this static global ranking risks utilizing stale difficulty assessments. It remains unclear whether this lag negatively impacts the training dynamics or if a more frequent, batch-level update would yield better pacing.
    • The claims of cross-organ generalizability are not sufficiently supported. The brain lesion detection experiment relies on a drastically simplified difficulty signal (whole-image gradients rather than anatomy-aware masking). The performance results in this domain are noticeably weaker and lack the rigorous statistical validation applied to the primary prostate task, undermining the assertion that the method is broadly applicable.
  • 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.

    (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 paper presents a clinically relevant approach to addressing case-level difficulty imbalance, supported by a novel anatomy-aware difficulty estimation strategy and strong empirical results on an independent test set. However, despite its practical value, there are several critical omissions. The framework lacks sufficient theoretical grounding for its underlying hard-to-easy strategy, and the absence of baseline comparisons against other established curriculum learning variants makes it difficult to assess the method’s true performance. The framework’s strict dependency on high-quality segmentation masks raises concerns about its real-world robustness, particularly without a corresponding sensitivity analysis. While the proposed method shows potential, these methodological and comparative gaps lead me to suggest a weak reject.

  • 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

    The paper proposes a hard-to-easy curriculum learning strategy (RSSPL) with anatomy-aware difficulty estimation to address case-level imbalance in prostate cancer detection. Extensive experiments demonstrate improved stability and performance, particularly at the patient level, with good generalizability across tasks.

  • 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 authors address the strong heterogeneity of PCa tumors by introducing a novel hard-to-easy curriculum learning strategy, providing an insightful reordering of training samples at the batch level. In the experimental section, they conduct comprehensive evaluations, including external validation, ablation studies, and cross-organ experiments, which together provide solid evidence for the effectiveness of the proposed method.

  • 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 authors restrict gradient computation to the prostate gland to reduce background noise. However, it remains unclear whether a more task-specific region, such as tumor region itself, could provide a precise signal. It is strongly recommended to include additional comparative experiments in table 3.2.The authors compare the performance of U-Net and MNet as backbones in Table 1.It would be beneficial to include more recent and advanced segmentation architectures to better demonstrate the superiority. 3.The authors clearly indicate in Fig. 1 that three input modalities are used (T2W, ADC, and DWI); however, the subsequent sections do not describe how these modalities are processed. Moreover, Fig. 2 only presents ADC and DWI images, without including T2W.

  • 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?

    This work lies in a ranking-based training order optimization strategy, and its effectiveness is preliminarily validated across different models and tasks. However, it lacks integration with or comparison against current state-of-the-art methods, which limits the strength of evidence for its superiority.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors fully answered my comments and provided convincing evidence.



Author Feedback

We sincerely thank the reviewers for the constructive feedback and positive recognition of the clinical motivation, model-agnostic design, and optimisation/stability benefits of Reverse Symmetric Self-Paced Learning (RSSPL). We address the concerns below. 1.Positioning and evaluation scope of RSSPL (R2&R3&R4&metaR): RSSPL is primarily proposed as a practical, model-agnostic optimisation strategy to enhance existing models rather than a theoretically driven curriculum learning framework or a new segmentation architecture. Conceptually, RSSPL extends hard-example-oriented optimisation from voxel-level weighting to case-level sample scheduling. a) Regarding backbone/architecture selection, since RSSPL is designed as a plug-and-play optimisation strategy, the primary objective was to evaluate RSSPL under different backbone and loss settings. nnUNet remains one of the strongest and most widely adopted baselines for small-data 3D segmentation-based detection, while MNet is a recent publicly available backbone validated as top-ranking in the PI-CAI challenge, outperforming the default UNet backbone within the nnUNet framework. Using both the default UNet and stronger MNet within the same nnUNet framework provides representative and controlled backbone settings for assessment. b) Regarding curriculum learning comparisons and observed gains, as shown in Table 1, EW denotes the conventional random training setting, while SSPL denotes the standard easy-to-hard curriculum learning paradigm. RSSPL was directly compared against both EW and SSPL. Compared with EW, RSSPL showed consistently improved performance, while compared with SSPL, RSSPL additionally demonstrated improved optimisation stability through consistently reduced cross-validation variance across all configurations, which is important in medical imaging settings. 2.Clarifying MRI processing details (R4&metaR). We used the public PI-CAI dataset together with the official PI-CAI processing tools, where ADC and highB were resampled to match the T2W voxel spacing and then adjusted to consistent spatial dimensions within each case. Subsequently, the standard nnUNet preprocessing command was executed for automatic planning, normalisation, and resampling, followed by standard nnUNet training procedures. We did not include the T2W in Fig. 2 primarily due to space, and because it is less important than ADC and highB for peripheral zone lesion assessment under PI-RADS guidelines. 3.Generalisability of RSSPL (R2&R3). The cross-organ experiment provides preliminary evidence of RSSPL transferability beyond prostate MRI. While prostate anatomy priors were used to constrain difficulty estimation in the primary task, this represents a prostate-specific implementation rather than a framework dependency. 4.Epoch-level ranking strategy and computational considerations (R2&R3). In 3D prostate MRI settings, GPU memory constraints limit training to batch size = 2, making batch-level ranking highly unstable and locally biased. Therefore, RSSPL estimates difficulty once per epoch as a practical stability-efficiency tradeoff. The measured training overhead fluctuated across epochs, with both slight increases and decreases observed, and no consistent increase. No inference overhead was introduced, as the inference pipeline remained unchanged. 5.Tumor-region vs. gland-based difficulty estimation (R3&R4). Tumor-region-based difficulty estimation is not directly applicable to negative cases, since no tumor regions are delineated for these samples. Lesion annotations are also generally more variable than gland segmentations, and strong lesion signals would naturally be reflected in gland-region signals. Therefore, prostate gland masks provide a more practical anatomical prior across both positive and negative cases. Since the gland masks are used as coarse anatomical priors rather than precise supervision, the proposed strategy is expected to be less sensitive to minor segmentation imperfections.




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.
    • Should include more up-to-date baseline methods for comparison.
    • Perform a complete literature review.
    • Provide details on how multiparametric MRI sequences were processed.
  • 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 answered reviewers’ comments and provided convincing evidence.



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 reviews agree that the paper addresses a clinically important and underexplored problem in prostate mpMRI analysis, namely, case-level difficulty imbalance during training. Reviewers consistently appreciated the practical and model-agnostic nature of the proposed Reverse Symmetric Self-Paced Learning (RSSPL) strategy, the clinically motivated anatomy-aware difficulty estimation, and the overall experimental design, including external validation and cross-task evaluation. In particular, reviewers highlighted that the method can be integrated into standard segmentation pipelines without architectural modifications and that the observed improvements in optimization stability and reproducibility are valuable for medical imaging applications.

    The main concerns raised during review related to (1) the limited theoretical grounding of the proposed hard-to-easy curriculum strategy, (2) the relatively modest magnitude of performance gains in some settings, (3) the dependence on prostate gland masks for anatomy-aware difficulty estimation, (4) the scope of the generalizability claims, and (5) the absence of comparisons against additional recent architectures or curriculum-learning variants.

    After considering the rebuttal and reviewer updates, I believe the authors addressed the major concerns sufficiently. In particular, the rebuttal successfully clarified the intended positioning of RSSPL as a practical optimization and sample-scheduling strategy rather than a theoretically driven curriculum learning framework or a new segmentation architecture. This clarification appropriately contextualizes the contribution and aligns the evaluation with the stated goals of the work.

    The rebuttal also provided convincing explanations regarding several methodological choices. The rationale for gland-based rather than lesion-based difficulty estimation was well justified, especially given the absence of lesion annotations for negative cases and the higher variability of lesion masks. The explanation of the epoch-level ranking strategy as a stability-efficiency tradeoff under small 3D MRI batch sizes was also reasonable, and the clarification that no inference overhead is introduced strengthens the practical relevance of the method. In addition, the authors appropriately moderated the claims regarding cross-organ generalizability and framed the secondary experiment as preliminary evidence of transferability rather than definitive proof of universal applicability.

    Importantly, the empirical evaluation remains solid overall. The paper includes experiments across multiple backbones and loss functions, external validation on Prostate158, statistical testing, ablation studies, and analyses of optimization stability. While the gains are incremental in some configurations, the improvements appear consistent across settings, particularly at the patient level, and the reduction in cross-validation variance is a meaningful contribution for clinically oriented AI systems.

    At the same time, several limitations remain and should be reflected carefully in the final version. The theoretical motivation for why hard-to-easy scheduling is preferable to conventional easy-to-hard curricula under polarized case distributions remains largely empirical rather than formally established. The cross-organ validation is still relatively limited, and the dependence on anatomical priors may affect robustness under imperfect gland segmentation. Furthermore, comparisons against additional recent training strategies or architectures could strengthen the broader positioning of the work.

    Overall, however, I find the paper to present a technically sound, practically useful, and clinically relevant contribution that is above the acceptance threshold. The rebuttal substantially improved the clarity and positioning of the work, and multiple reviewers updated their recommendations positively after the rebuttal.



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.

    Congratulations!



back to top