List of Papers Browse by Subject Areas Author List
Abstract
Micro-ultrasound (μUS) is a new, emerging, and promising imaging modality for prostate cancer (PCa) detection, but accurate identification of suspicious tissue remains highly dependent on clinical experience, leading to substantial inter-observer variability. Machine-learning assistance can reduce this variability; however, training reliable deep models is challenging because supervision is sparse and noisy—typically limited to core-level histopathology outcomes (e.g., cancer grade and its percentage in a biopsy core) without pixel-level lesion annotations and under severe class imbalance. We introduce Prost-RL, which reframes μUS PCa detection as a spatially aware, policy-driven inference problem by learning where to look before decoding. Prost-RL integrates a lightweight reinforcement-learning policy into a foundation-model encoder–decoder to generate interpretable spatial attention maps that act as soft prompts for both cancer-likelihood heatmap prediction and image-level classification. We further propose Adaptive Policy Optimization (APO) to stabilize hybrid supervised–RL training and a noise-robust objective combining symmetric cross-entropy with negative-entropy regularization to mitigate weak-label noise and encourage sharp localization. On a cohort of 6,607 biopsy cores from 693 patients across five clinical sites, Prost-RL achieves 79.0 ± 3.5 AUROC with 64.6 ± 6.3% sensitivity at 80% specificity for core-level detection (+2.1 AUROC and +4.5 sensitivity points over the strongest baseline), and 79.3 ± 5.8 AUROC for clinically significant cancer classification. The learned policy highlights biopsy-aligned regions, providing transparent, spatially grounded evidence alongside quantitative risk predictions. Code is available at: https://github.com/DeepRCL/Prost-RL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6168_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/DeepRCL/Prost-RL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{AboMoh_Learning_MICCAI2026,
author = { Abootorabi, Mohammad Mahdi AND Namazi, Sina AND Saadat, Armin AND Wang, Lyuyang AND Dzikunu, Obed K. AND Wilson, Paul F. R. AND Guo, Zhuoxin AND Wodlinger, Brian AND Mousavi, Parvin AND Abolmaesumi, Purang},
title = { { Learning Where to Look: A Reinforcement Learning Framework for Robust Micro-ultrasound 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 addresses micro-ultrasound prostate cancer detection under core-level weak supervision, without pixel-level annotations, and with severe class imbalance. Built on ProstNFound+, it proposes Prost-RL, which adds an explicit spatial policy network, noise-robust losses, and RL fine-tuning to learn where to attend before decoding, aiming to improve both heatmap localization and core-level / csPCa classification.
- 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)It studies a clinically important problem: how to detect prostate cancer under biopsy core-level weak supervision. (2)It proposes Prost-RL and shows performance gains in both image-level csPCa classification and core-level detection.
- 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)Please report the results of statistical significance tests. (2) The spatial attention mechanism and downstream feature modulation network are end-to-end differentiable. The pairwise ranking reward in Eq. (1) could likely be optimized with a standard smooth ranking loss through backpropagation alone. Please justify why RL is necessary.
- 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?
please refer to weakness
- 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
The authors propose an iterative improvement over a previous method for prostate cancer detection on microultrasound, which consists of a reinforcement learning-based tuning stage and techniques to reduce the impact of data imbalance.
- 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.
Strong validation, with cross-validation on multi-center data Demonstrated performance improvement over previous methods Ablation studies show the impact of each component
- 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 marginal improvements over ProstNFound+ might not justify the added complexity to the method.
Is the ground truth compared to other grading methods, such as targeted biopsy or radical prostatectomy? As needle biopsy can have 30-50% difference from radical prostatectomy grading, I wonder if this was validated in this dataset.
- 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 provide sufficient information for 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?
Sound methodology, but I’m not sure if the added complexity is justified for the marginal performance gains (<1% AUC for csPCa)
- 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
Review #3
- Please describe the contribution of the paper
The paper proposed a RL approach to predict for areas of high likelihood for prostate cancer on biopsies.
- 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 showed the feasibility to use a novel RL algorithm (DRPO) for the task of prostate cancer detection. The experiments were well designed.
- 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 impact might be limited, since this method is built on top of the ProstNFound+ method, a paper that was published in Feb 2026 and has no citation so far.
- 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 provide sufficient information for 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?
Well carried out experiments and moderately novel idea, but limited impact.
- 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
We sincerely thank the reviewers and meta-reviewer for recommending our paper for provisional acceptance and for highlighting its clinical relevance, well-designed multi-center cross-validation, consistent performance gains over strong baselines, informative ablation studies, and sound methodology. We further clarify a few points below.
1.Biopsy vs. Prostatectomy labels (R3, meta-reviewer). R3 raises an important clinical point: needle biopsy samples only a fraction of the gland and can disagree with whole-mount prostatectomy grading, making core-level labels intrinsically noisy. This is precisely the regime our noise-robust design (SCE plus entropy regularization) is built to handle. Since prostatectomy outcomes are not systematically available for multi-center µUS cohorts, we will acknowledge this in the Discussion and identify validation on a surgical sub-cohort as future work.
2.Magnitude and significance of gains, and generality of the contribution (R2, R3, R4, meta-reviewer). In our setting with ~86% benign cores and noisy core-level labels, the gains concentrate where they matter clinically: the improvements in core-level AUROC and in sensitivity at 80% specificity (+4.5 points over the strongest baseline) are statistically significant, with consistent positive trends across the remaining metrics. The margin also widens at lower tumor-involvement thresholds (Fig. 2a), i.e., on the subtle lesions most often missed. We will report the statistical-significance details alongside the existing results in the camera-ready version. Qualitatively (Fig. 3), Prost-RL yields sharper, biopsy-aligned heatmaps and an interpretable policy attention map supporting clinician trust independently of AUROC. Beyond performance, our contributions (the spatial attention policy, the SCE plus entropy objective, and APO) are encoder–decoder agnostic; we instantiate them on ProstNFound+ as the current SOTA, but the same recipe applies to other weakly supervised medical-imaging tasks with sparse, noisy, proportion-based labels and class imbalance, which we plan to explore in future work.
3.Why RL is necessary, vs. a differentiable ranking surrogate (R2). R2 correctly notes that Eq. (1) could be optimized with a smooth ranking surrogate. However, RL offers two properties that direct backpropagation lacks in this setting:
(i) Exploration of attention configurations: Backpropagation refines the current attention map along its local gradient. Our policy instead generates K stochastic rollouts per image, observes how different spatial focal patterns affect downstream ranking, and updates toward configurations empirically associated with better separation. This is an exploration mechanism, not only a gradient direction.
(ii) Adaptive reweighting from rollout statistics: With ~86% benign cores and wide variance in case difficulty (e.g., low- vs. high-involvement tumors), easy cases dominate a per-sample gradient. Each sample in our setup yields a vector of K rewards from its K rollouts, summarizing how the model is currently performing on that sample. Our method clusters samples whose reward vectors are similar, separately within cancer and benign groups, then up-weights small (rare) or poorly performing (hard) clusters through hierarchical temperature scaling. Because supervised training produces a single deterministic loss per sample rather than a reward vector, this clustering has no backpropagation analog.
Our ablation in Fig. 2(a–b) provides empirical support for these two ingredients: the RL stage produces its largest margins over its supervised counterpart at low tumor-involvement thresholds, the regime where ranking subtle lesions matters most, indicating that exploration and adaptive reweighting contribute gains beyond what standard supervised gradients alone can deliver.
We again thank the reviewers and meta-reviewer for their feedback.
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 three reviewers rate the paper as Weak Accept, and I agree with this assessment and recommend acceptance.
The paper addresses a clinically relevant problem and demonstrates good results. Please address the following concerns in the final version: add a statistical significance test, and add to teh discussion the ground-truth reliability (biopsy vs. prostatectomy).