{"id":"b5f3edf9-295a-4adf-8885-f243d93768be","arxiv_id":"2606.30951","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Prost-RL integrates an RL policy into a foundation-model encoder-decoder to generate interpretable spatial attention maps that improve core-level prostate cancer detection in micro-ultrasound, achieving 79.0 AUROC on a 6,607-core multi-site dataset.","lead":"The paper introduces Prost-RL, a reinforcement learning framework that learns spatial attention policies to guide micro-ultrasound analysis for prostate cancer detection under weak and noisy core-level labels. A smart generalist might read it to understand how policy-driven attention can improve consistency in medical imaging AI when pixel-level annotations are unavailable.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Multi-site performance claim rests on unverified assumption that pooled training generalizes across sites rather than capturing site-specific artifacts","rationale":"The reader's weakest assumption (stable RL training and generalization without overfitting to site-specific artifacts) is exactly the load-bearing point for the multi-site empirical claim. No additional internal inconsistency is detectable from the abstract; the concern is therefore already captured and the CONDITIONAL verdict with LOW confidence remains appropriate.","tokens_in":1904,"tokens_out":333,"duration_ms":26252,"concrete_test":"Re-train and evaluate using leave-one-site-out cross-validation across the five sites; report mean AUROC and sensitivity at 80% specificity on each held-out site. If any held-out site shows >5-point AUROC drop relative to the pooled figure or the improvement over the strongest baseline vanishes, the multi-site generalization claim does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result (79.0 AUROC, +2.1 over baseline on 6,607 cores from 5 sites) requires that the RL policy plus APO training produces features robust to domain shift. The abstract states the cohort spans five clinical sites but supplies no information on split strategy (patient-level, site-stratified, or leave-one-site-out). In ultrasound, probe frequency response, gain settings, and labeling conventions routinely differ by site; a random pooled split can therefore inflate metrics by allowing the attention maps to exploit site cues while still appearing to improve over baselines. This directly weakens the claim that the observed gains are attributable to the spatially-aware RL formulation rather than dataset artifacts.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","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.","tokens_in":2061,"tokens_out":547,"duration_ms":45217,"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":[{"comment":"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.","section":"Abstract"},{"comment":"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.","section":"Abstract"}],"minor_comments":[{"comment":"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.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"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.","responses":[{"response":"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_made":"yes","referee_comment":"[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."},{"response":"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_made":"yes","referee_comment":"[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."}],"tokens_in":1614,"tokens_out":418,"duration_ms":41080,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main point is that Prost-RL trains a lightweight RL policy to produce soft spatial prompts inside a foundation-model encoder-decoder for core-level prostate cancer detection in micro-ultrasound. It pairs this with Adaptive Policy Optimization and a symmetric cross-entropy plus negative-entropy loss to handle noisy weak labels. On 6607 cores from 693 patients across five sites it reports 79.0±3.5 AUROC and a 2.1-point gain over the strongest baseline, plus public code.\n\nThe combination of RL-driven attention with that specific loss for noisy supervision is not a direct copy of prior attention or supervised baselines in the cited work. The multi-site scale and standard deviations give the empirical numbers more weight than single-center studies usually carry.\n\nThe clearest gap is the missing description of the train-test split. The abstract says the data come from five sites but does not state whether the split is patient-level, site-stratified, or leave-one-site-out. In μUS, probe settings and labeling habits differ by site, so a pooled random split can let the policy exploit site cues while still beating baselines. That directly affects how much credit the RL formulation deserves for the reported robustness. Ablations are also absent from the abstract, so it is hard to isolate the contribution of the policy versus the foundation model or the loss terms.\n\nThe paper is aimed at groups working on weak-supervision methods in ultrasound or prostate imaging. A reader who needs a concrete RL-plus-foundation-model recipe for attention under core-level labels will find usable details and code.\n\nIt deserves peer review. The dataset size and public code make the work worth a referee's time to check the split and ablations, even if the gains stay modest after that check.","headline":"Prost-RL gets a modest AUROC lift on a 5-site μUS cohort by adding an RL policy for spatial attention, but the multi-site claim hinges on an unshown split strategy.","tokens_in":2614,"tokens_out":443,"would_cite":false,"duration_ms":8925,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Prost-RL trains a reinforcement learning policy to generate spatial attention maps that guide prostate cancer detection in micro-ultrasound images.","keywords":["reinforcement learning","prostate cancer detection","micro-ultrasound","attention maps","medical image analysis","weak supervision","multi-site validation"],"falsifier":"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.","tokens_in":2809,"feed_emoji":"🔬","tokens_out":670,"duration_ms":23995,"temperature":0.7,"pith_summary":"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.","feed_headline":"RL policy learns to focus attention for prostate cancer detection in micro-ultrasound","feed_subtitle":"Prost-RL reaches 79.0 AUROC on 6607 cores from five sites by generating spatial prompts before classification.","key_machinery":"The reinforcement-learning policy that produces spatial attention maps as soft prompts for the decoder.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["RL learns where to look before decoding micro-ultrasound prostate cancer","Prost-RL applies lightweight RL for spatial attention in μUS PCa detection","Attention maps generated by RL policy for micro-ultrasound cancer detection","RL policy provides spatial prompts for PCa detection in μUS imaging"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["RL learns where to look before decoding micro-ultrasound prostate cancer","Prost-RL applies lightweight RL for spatial attention in μUS PCa detection","Attention maps generated by RL policy for micro-ultrasound cancer detection","RL policy provides spatial prompts for PCa detection in μUS imaging"]},"model":"grok-4.3","cost_usd":0.007937,"raw_usage":{"total_tokens":3679,"prompt_tokens":793,"num_sources_used":0,"completion_tokens":74,"cost_in_usd_ticks":79374500,"prompt_tokens_details":{"text_tokens":793,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2812,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":793,"tokens_out":74,"duration_ms":34848,"temperature":1.0,"reasoning_tokens":2812,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-01T01:31:44.080922+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"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.","supporting_citations":[],"review_version":1}