A research team has built a reinforcement-learning system that decides where to examine a micro-ultrasound image before attempting to detect prostate cancer.
The model, called Prost-RL, tackles a specific bottleneck in AI-assisted biopsy analysis: training data is weak. Pathology labs record outcomes at the level of an entire biopsy core — cancer grade, percentage of involvement — not pixel-by-pixel lesion maps. That sparse supervision makes it hard to teach a neural network exactly where to look. Prost-RL sidesteps this by learning a spatial attention policy first, then using those attention maps as soft prompts for both a cancer-likelihood heatmap and an image-level classification. Tested on 6,607 biopsy cores from 693 patients across five clinical sites, it reached 79.0 AUROC with 64.6% sensitivity at 80% specificity, beating the next-best baseline by 2.1 AUROC points and 4.5 sensitivity points.
The real gap Prost-RL is closing is inter-observer variability — experienced radiologists and urologists read micro-ultrasound differently, and that inconsistency is the reason many centers still default to conventional MRI. A system that generates interpretable spatial attention maps alongside a risk score gives clinicians something to interrogate, not just a black-box answer. Micro-ultrasound is also cheaper and more portable than MRI, so if AI assistance can make it reliable, it widens access.
It is worth noting that 79 AUROC, while a meaningful step up, still leaves substantial room before any tool earns a role in routine clinical decisions — and a five-site cohort, while respectable for academic work, is not a regulatory submission.