pith. sign in

arxiv: 2606.30951 · v1 · pith:BGVRDRWInew · submitted 2026-06-29 · 💻 cs.CV · cs.AI· cs.LG

Learning Where to Look: A Reinforcement Learning Framework for Robust Micro-Ultrasound Prostate Cancer Detection

Mohammad Mahdi Abootorabi , Sina Namazi , Armin Saadat , Lyuyang Wang , Obed Dzikunu , Paul F. R. Wilson , Zhuoxin Guo , Brian Wodlinger
show 2 more authors
Parvin Mousavi Purang Abolmaesumi
This is my paper

Pith reviewed 2026-07-01 01:31 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LG
keywords reinforcement learningprostate cancer detectionmicro-ultrasoundattention mapsmedical image analysisweak supervisionmulti-site validation
0
0 comments X

The pith

Prost-RL trains a reinforcement learning policy to generate spatial attention maps that guide prostate cancer detection in micro-ultrasound images.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents Prost-RL as a method that treats micro-ultrasound prostate cancer detection as a policy-driven task where the model first learns where to focus before predicting cancer likelihood. It combines a lightweight RL policy with a foundation-model encoder-decoder, uses Adaptive Policy Optimization for stable hybrid training, and adds a noise-robust loss to handle sparse core-level labels. On 6607 biopsy cores from 693 patients at five sites, the approach reports 79.0 AUROC for core-level detection and 79.3 AUROC for clinically significant cancer, with gains over the strongest baseline. The attention maps are positioned as interpretable outputs that align with biopsy regions.

Core claim

Prost-RL reframes μUS PCa detection as a spatially aware, policy-driven inference problem by learning where to look before decoding, integrating 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, stabilized by Adaptive Policy Optimization and a noise-robust objective combining symmetric cross-entropy with negative-entropy regularization.

What carries the argument

The reinforcement-learning policy that produces spatial attention maps as soft prompts for the decoder.

If this is right

  • The learned attention maps supply spatially grounded evidence alongside quantitative risk scores.
  • Hybrid supervised-RL training with APO can stabilize learning under weak supervision and class imbalance.
  • Core-level detection reaches 79.0 AUROC and 64.6% sensitivity at 80% specificity, with a 2.1 AUROC gain over the strongest baseline.
  • Clinically significant cancer classification reaches 79.3 AUROC.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same policy-learning step could be tested on other ultrasound or MRI tasks that lack pixel-level annotations.
  • If attention maps remain consistent across new clinical sites, the method could support standardized reading workflows that reduce inter-observer differences.
  • Real-time inference speed of the lightweight policy would determine whether the approach fits inside existing biopsy procedures.

Load-bearing premise

A reinforcement learning policy trained only on core-level histopathology labels can produce attention maps that generalize across sites without overfitting to label noise or imaging artifacts.

What would settle it

Retraining the policy on the same multi-site data and finding that the resulting attention maps fail to highlight biopsy-aligned regions or that AUROC on held-out sites falls below the reported baseline.

Figures

Figures reproduced from arXiv: 2606.30951 by Armin Saadat, Brian Wodlinger, Lyuyang Wang, Mohammad Mahdi Abootorabi, Obed Dzikunu, Parvin Mousavi, Paul F. R. Wilson, Purang Abolmaesumi, Sina Namazi, Zhuoxin Guo.

Figure 1
Figure 1. Figure 1: Prost-RL overview. Image and clinical prompt encoders produce features F and embedding c. A spatial policy πθ generates an attention map α that modulates F into shared embeddings E, decoded by a heatmap head and a csPCa classifier. Canada), acquiring B-mode sagittal-plane images (depth 28 mm, width 46.06 mm) at 10–12 cores per patient; the frame immediately preceding needle firing was extracted per core. C… view at source ↗
Figure 2
Figure 2. Figure 2: Top: comparison of Prost-RL (Full) vs. supervised-only variant and other base￾lines. Bottom: ablation study results for different setting choices. with sensitivity at 80% specificity improving from 60.1% to 64.6%. Gains extend to high-involvement cores (84.9 vs. 83.6) and csPCa detection (79.9 vs. 79.6). MedSAM-based approaches without explicit spatial guidance lag considerably behind, underscoring the ben… view at source ↗
Figure 3
Figure 3. Figure 3: Example heatmaps generated by ProstRL and ProstNFound+, shown alongside ground-truth labels and radiologist annotations. improves the model’s ability to rank and localize subtle, low-involvement lesions that supervised training alone struggles to distinguish. Ablation Studies and Qualitative Evaluation [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
read the original abstract

Micro-ultrasound ($\mu$US) 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 $\mu$US 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 $79.0\pm3.5$ AUROC with $64.6\pm6.3$% sensitivity at 80% specificity for core-level detection (+2.1 AUROC and +4.5 sensitivity points over the strongest baseline), and $79.3\pm5.8$ 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces Prost-RL, a reinforcement-learning framework that integrates a policy network into a foundation-model encoder-decoder to produce spatial attention maps for micro-ultrasound prostate cancer detection. These maps serve as soft prompts for core-level cancer-likelihood heatmaps and image-level classification. Training uses Adaptive Policy Optimization (APO) together with a symmetric cross-entropy plus negative-entropy regularization objective to handle sparse, noisy core-level histopathology labels. On a cohort of 6,607 biopsy cores from 693 patients across five clinical sites, the method reports 79.0±3.5 AUROC and 64.6±6.3% sensitivity at 80% specificity for core-level detection (+2.1 AUROC and +4.5 sensitivity over the strongest baseline) and 79.3±5.8 AUROC for clinically significant cancer classification, with public code released.

Significance. If the reported gains are shown to arise from the spatially-aware RL policy rather than site-specific cues, the work would supply a practical route to interpretable, weakly-supervised detection in μUS imaging. Notable strengths include the multi-site cohort size, reported standard deviations, patient-level held-out evaluation against external baselines, and public code release.

major comments (2)
  1. [Abstract] Abstract: the headline multi-site claim (79.0±3.5 AUROC across five clinical sites) is load-bearing on the assumption that the learned attention policy generalizes across sites rather than exploiting site-specific ultrasound artifacts (probe frequency, gain, labeling conventions). No information is supplied on the data partitioning strategy (patient-level, site-stratified, or leave-one-site-out), which is required to substantiate that the +2.1 AUROC improvement is attributable to the RL formulation.
  2. [Abstract] Abstract (hybrid training description): the central claim that APO stabilizes training of the RL policy on core-level labels alone rests on the unverified assumption that the resulting attention maps do not overfit to label noise or site artifacts; the abstract supplies neither ablation results on the loss-component weights nor sensitivity analysis of APO hyperparameters, leaving the robustness of the 79.0 AUROC figure only moderately anchored.
minor comments (1)
  1. [Abstract] Abstract: the phrase 'foundation-model encoder-decoder' is introduced without naming the specific backbone or its pre-training corpus, which affects reproducibility of the reported numbers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and outline the revisions that will be made to clarify the evaluation protocol and strengthen the supporting analyses.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline multi-site claim (79.0±3.5 AUROC across five clinical sites) is load-bearing on the assumption that the learned attention policy generalizes across sites rather than exploiting site-specific ultrasound artifacts (probe frequency, gain, labeling conventions). No information is supplied on the data partitioning strategy (patient-level, site-stratified, or leave-one-site-out), which is required to substantiate that the +2.1 AUROC improvement is attributable to the RL formulation.

    Authors: We agree that the data partitioning strategy is essential to substantiate the multi-site generalization claim. The current manuscript does not supply this information. We will revise the methods section to explicitly describe the partitioning as patient-level with site-stratified sampling and will add leave-one-site-out experiments to demonstrate that the reported gains arise from the RL policy rather than site-specific cues. revision: yes

  2. Referee: [Abstract] Abstract (hybrid training description): the central claim that APO stabilizes training of the RL policy on core-level labels alone rests on the unverified assumption that the resulting attention maps do not overfit to label noise or site artifacts; the abstract supplies neither ablation results on the loss-component weights nor sensitivity analysis of APO hyperparameters, leaving the robustness of the 79.0 AUROC figure only moderately anchored.

    Authors: We acknowledge that the manuscript does not currently include ablations on loss-component weights or sensitivity analysis of APO hyperparameters. We will add these analyses (varying the symmetric cross-entropy and negative-entropy weights as well as APO hyperparameters) to the revised manuscript or supplementary material to better anchor the robustness of the training procedure. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical claims rest on held-out multi-site evaluation

full rationale

The paper describes a hybrid RL-supervised model (Prost-RL with APO) trained on core-level labels and reports AUROC/sensitivity on a held-out cohort of 6,607 cores from 693 patients across five sites, with explicit gains over external baselines. No equations, self-citations, or fitted parameters are presented that reduce the headline metrics or attention-map claims to quantities defined inside the paper by construction; the derivation chain remains independent of its own outputs.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim depends on standard deep-learning assumptions plus two domain-specific premises about RL policy learning from weak labels; no new physical entities are postulated and the number of explicitly free parameters is modest but not fully enumerated in the abstract.

free parameters (2)
  • loss-component weights for symmetric cross-entropy and negative-entropy regularization
    Balance terms in the noise-robust objective; values chosen during training but not reported.
  • APO and RL policy hyperparameters
    Learning rates, reward scaling, and policy network size required for stable hybrid training.
axioms (2)
  • domain assumption A foundation-model encoder-decoder can be effectively conditioned by soft spatial attention maps generated by an RL policy.
    Invoked when the policy output is described as acting as soft prompts for heatmap and classification heads.
  • domain assumption Core-level histopathology labels supply sufficient supervisory signal for an RL policy to learn biopsy-aligned spatial attention without pixel-level ground truth.
    Central premise of the reframing as a spatially aware policy-driven inference problem.

pith-pipeline@v0.9.1-grok · 5902 in / 1643 out tokens · 68715 ms · 2026-07-01T01:31:44.080922+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    IEEE Access11, 118217–118228 (2023)

    Alzate-Grisales, J.A., Mora-Rubio, A., García-García, F., Tabares-Soto, R., De La Iglesia-Vayá, M.: Sam-unetr: Clinically significant prostate cancer segmentation using transfer learning from large model. IEEE Access11, 118217–118228 (2023)

    2023

  2. [2]

    IEEE Access11, 118217–118228 (2023)

    Alzate-Grisales, J.A., Mora-Rubio, A., García-García, F., Tabares-Soto, R., De La Iglesia-Vayá, M.: Sam-unetr: Clinically significant prostate cancer segmentation using transfer learning from large model. IEEE Access11, 118217–118228 (2023). https://doi.org/10.1109/ACCESS.2023.3326882

    work page doi:10.1109/access.2023.3326882 2023

  3. [3]

    CA: a cancer journal for clinicians pp

    Bray, F., Laversanne, M., Sung, H., Ferlay, J., Siegel, R.L., Soerjomataram, I., Je- mal, A.: Global cancer statistics 2022: Globocan estimates of incidence and mor- tality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians pp. 229–263 (2024)

    2022

  4. [4]

    arXiv preprint arXiv:2506.00711 (2025)

    Dai, W., Chen, P., Ekbote, C., Liang, P.P.: Qoq-med: Building multimodal clinical foundation models with domain-aware grpo training. arXiv preprint arXiv:2506.00711 (2025)

    work page arXiv 2025

  5. [5]

    Tomography11(7), 80 (2025)

    DuBois, J., Smani, S., Golos, A., Rivera Lopez, C., Lokeshwar, S.D.: Micro- ultrasound in the detection of clinically significant prostate cancer: A compre- hensive review and comparison with multiparametric mri. Tomography11(7), 80 (2025)

    2025

  6. [6]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Elghareb, T., Harmanani, M., To, M.N.N., Wilson, P., Jamzad, A., Fooladgar, F., Abdelsamad, B., Dzikunu, O., Sojoudi, S., Reznik, G., et al.: Proteus: A spatio- temporal enhanced ultrasound-based framework for prostate cancer detection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 405–414. Springer (2025)

    2025

  7. [7]

    Computers12(8), 152 (2023)

    Gavade, A.B., Nerli, R., Kanwal, N., Gavade, P.A., Pol, S.S., Rizvi, S.T.H.: Auto- mated diagnosis of prostate cancer using mpmri images: A deep learning approach for clinical decision support. Computers12(8), 152 (2023)

    2023

  8. [8]

    In: International conference on medical image computing and computer-assisted intervention

    Gilany, M., Wilson, P., Jamzad, A., Fooladgar, F., To, M.N.N., Wodlinger, B., Abolmaesumi, P., Mousavi, P.: Towards confident detection of prostate cancer using high resolution micro-ultrasound. In: International conference on medical image computing and computer-assisted intervention. pp. 411–420. Springer (2022)

    2022

  9. [9]

    IEEE Transactions on Medical Imaging (2025)

    Han, K., Wang, S., Chen, J., Qian, C., Lyu, C., Ma, S., Qiu, C., Sheng, V.S., Huang, Q., Liu, Z.: Region uncertainty estimation for medical image segmentation with noisy labels. IEEE Transactions on Medical Imaging (2025)

    2025

  10. [10]

    In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI)

    Harmanani, M., Jamzad, A., To, M.N.N., Wilson, P.F., Guo, Z., Fooladgar, F., Sojoudi, S., Gilany, M., Chang, S., Black, P., et al.: Cinepro: Robust training of foundation models for cancer detection in prostate ultrasound cineloops. In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI). pp. 1–5. IEEE (2025)

    2025

  11. [11]

    Harmanani, M., Wilson, P.F., To, M.N.N., Gilany, M., Jamzad, A., Fooladgar, F., Wodlinger, B., Abolmaesumi, P., Mousavi, P.: Trusworthy: toward clinically appli- cable deep learning for confident detection of prostate cancer in micro-ultrasound. 10 M. M. Abootorabi et al. International Journal of Computer Assisted Radiology and Surgery20(5), 981–989 (2025)

    2025

  12. [12]

    Translational andrology and urol- ogy6(3), 345 (2017)

    Hutchinson, R., Lotan, Y.: Cost consideration in utilization of multiparametric magnetic resonance imaging in prostate cancer. Translational andrology and urol- ogy6(3), 345 (2017)

    2017

  13. [13]

    Computerized Medical Imaging and Graphics112, 102326 (2024)

    Jiang, H., Imran, M., Muralidharan, P., Patel, A., Pensa, J., Liang, M., Beni- dir, T., Grajo, J.R., Joseph, J.P., Terry, R., et al.: Microsegnet: A deep learning approach for prostate segmentation on micro-ultrasound images. Computerized Medical Imaging and Graphics112, 102326 (2024)

    2024

  14. [14]

    Canadian Urological Association Journal15(1), E11 (2020)

    Klotz, L., Lughezzani, G., Maffei, D., Sánchez, A., Pereira, J.G., Staerman, F., Cash, H., Luger, F., Lopez, L., Sanchez-Salas, R., et al.: Comparison of micro- ultrasound and multiparametric magnetic resonance imaging for prostate cancer: a multicenter, prospective analysis. Canadian Urological Association Journal15(1), E11 (2020)

    2020

  15. [15]

    Nature Communications15, 654 (2024)

    Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications15, 654 (2024)

    2024

  16. [16]

    Nature communications15(1), 654 (2024)

    Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature communications15(1), 654 (2024)

    2024

  17. [17]

    European Radiology28(3), 1016–1026 (2018)

    Reisæter, L.A., Fütterer, J.J., Losnegård, A., Nygård, Y., Monssen, J., Gravdal, K., Halvorsen, O.J., Akslen, L.A., Biermann, M., Haukaas, S., et al.: Optimising preoperative risk stratification tools for prostate cancer using mpmri. European Radiology28(3), 1016–1026 (2018)

    2018

  18. [18]

    Ultrasound in medicine & biology44(7), 1341–1354 (2018)

    Rohrbach, D., Wodlinger, B., Wen, J., Mamou, J., Feleppa, E.: High-frequency quantitative ultrasound for imaging prostate cancer using a novel micro-ultrasound scanner. Ultrasound in medicine & biology44(7), 1341–1354 (2018)

    2018

  19. [19]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv. org/abs/2402.033002(3), 5 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    IEEE Transactions on Medical Imaging (2025)

    Wang, H., Huai, L., Li, W., Qi, L., Jiang, X., Shi, Y.: Weakmedsam: Weakly- supervised medical image segmentation via sam with sub-class exploration and prompt affinity mining. IEEE Transactions on Medical Imaging (2025)

    2025

  21. [21]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 322–330 (2019)

    2019

  22. [22]

    Willis, E., Elghareb, T., Wilson, P.F.R., To, M.N.N., Abootorabi, M.M., Jamzad, A., Wodlinger, B., Mousavi, P., Abolmaesumi, P.: Guide-us: Grade-informed un- paired distillation of encoder knowledge from histopathology to micro-ultrasound (2026), https://arxiv.org/abs/2602.19005

    work page arXiv 2026

  23. [23]

    International Journal of Computer Assisted Radiology and Surgery pp

    Wilson, P.F., Harmanani, M., To, M.N.N., Jamzad, A., Elghareb, T., Guo, Z., Kinnaird, A., Wodlinger, B., Abolmaesumi, P., Mousavi, P.: Prostnfound+: A prospective study using medical foundation models for prostate cancer detection. International Journal of Computer Assisted Radiology and Surgery pp. 1–10 (2026)

    2026

  24. [24]

    Wilson, P.F., To, M.N.N., Jamzad, A., Gilany, M., Harmanani, M., Elghareb, T., Fooladgar, F., Wodlinger, B., Abolmaesumi, P., Mousavi, P.: Prostnfound: Inte- grating foundation models with ultrasound domain knowledge and clinical context forrobustprostatecancerdetection.In:InternationalConferenceonMedicalImage Computing and Computer-Assisted Intervention....

    2024