This study prospectively validated a semi-supervised learning framework using a lesion-only teacher model, referred to as RG-SSL-LOC, for scalable detection of clinically significant prostate cancer on biparametric MRI (bpMRI). The work evaluated whether a report-guided semi-supervised approach focused on lesions could improve segmentation quality and case-level diagnostic performance compared with fully supervised training and a prior report-guided semi-supervised method (RG-SSL). The study also assessed the added value of combining model predictions with clinico-radiologic variables in multimodal models.
Model development and validation used a multicenter dataset totaling 13,706 bpMRI examinations from 13,630 patients acquired across 27 centers. The dataset sources included publicly available repositories (PI-CAI, UCLA), the PROMIS dataset (available on request), and the ProstateNET dataset collected within the ProCAncer-I project (not publicly available due to privacy and consortium agreements). The authors report that appropriate ethics committee approvals, informed consent for prospective data, and waivers for retrospective data were obtained at each contributing site.
Three segmentation models were evaluated:
Comparative assessment was performed at both lesion and case levels across external retrospective, external prospective, and internal prospective cohorts.
Performance was measured at the lesion level using segmentation overlap metrics (median Dice) and at the case level using area under the receiver operating characteristic curve (AUC). Case-level results were compared to PI-RADS where applicable. Statistical comparisons and p-values are reported for pairwise model comparisons and for multimodal integration versus clinico-radiologic variables alone.
At the lesion level, the RG-SSL-LOC model achieved a higher median Dice score than both comparison methods. Reported median Dice values were 0.49 for RG-SSL-LOC, 0.41 for FSL, and 0.40 for RG-SSL. The differences in median Dice of RG-SSL-LOC versus FSL and versus RG-SSL were both statistically significant (p<.001), indicating improved segmentation quality for the lesion-focused semi-supervised teacher model.
RG-SSL-LOC achieved case-level AUCs of:
Comparisons versus FSL produced AUC differences with p-values reported as 0.237 (external retrospective), 0.020 (external prospective), and <.001 (internal prospective). Comparisons versus RG-SSL reported p-values of 0.929 (external retrospective), 0.652 (external prospective), and 0.007 (internal prospective). When compared with PI-RADS, reported AUCs and p-values were 0.78 (p=.055), 0.83 (p=.652), and 0.86 (p=.480) for the respective cohorts. These results indicate variable statistical significance across cohorts and comparisons.
Predictions from the best-performing model (RG-SSL-LOC) were combined with clinico-radiologic variables to form multimodal models. Multimodal integration produced significant improvements in AUC compared with clinico-radiologic variables alone in all three cohorts reported:
These findings indicate that adding RG-SSL-LOC predictions to routine clinical and radiologic variables improved discrimination in each tested cohort.
In the internal prospective cohort, the multimodal approach that included RG-SSL-LOC predictions reduced unnecessary biopsies by 15.19% compared with clinico-radiologic variables alone. The authors present this as evidence that the model may contribute to more selective biopsy decisions when integrated into multimodal decision tools.
The report reiterates that some datasets are publicly available (PI-CAI, UCLA), PROMIS is available on reasonable request, and ProstateNET is restricted by consortium agreements and privacy considerations. The work was supported by the European Union Horizon 2020 research and innovation program (grant 952159).
The article is a medRxiv preprint and has not undergone peer review; the authors and the preprint platform caution that the findings should not yet be used to guide clinical practice. The manuscript declares no competing interests. Specific methodological details, model training hyperparameters, and other implementation-level information are referenced in the full preprint and supplementary materials; if not reported in the source abstract, those details were not available in the provided source text.
In this multicenter prospective validation, the lesion-only report-guided semi-supervised model (RG-SSL-LOC) produced superior lesion segmentation (Dice) compared with fully supervised and prior semi-supervised methods and yielded robust case-level discrimination across external and internal cohorts. Integration of RG-SSL-LOC outputs with clinico-radiologic variables significantly increased AUC in all tested cohorts and reduced unnecessary biopsies in the internal prospective cohort, supporting the potential utility of lesion-focused semi-supervised learning for scalable prostate cancer detection on bpMRI. As a preprint, results require peer review and further external validation before clinical adoption.