Abstract
<jats:title>Abstract</jats:title> <jats:sec> <jats:title>Purpose</jats:title> <jats:p>To prospectively validate a semi-supervised learning framework with a lesion-only teacher model (RG-SSL-LOC) for scalable clinically significant prostate cancer detection on biparametric MRI (bpMRI) and assess its added value in multimodal models.</jats:p> </jats:sec> <jats:sec> <jats:title>Materials and Methods</jats:title> <jats:p>A multicenter dataset of 13,706 bpMRI examinations (13,630 patients, 27 centers) was used for model development/validation. Three segmentation models (fully supervised learning [FSL], a state-of-the-art report-guided semi-supervised approach [RG-SSL], and the proposed RG-SSL-LOC) were evaluated at lesion- and case-level on external retrospective, external prospective, and internal prospective cohorts. Predictions from the best-performing model were combined with clinico-radiologic variables in a multimodal approach. All case-level results were compared with PI-RADS.</jats:p> </jats:sec> <jats:sec> <jats:title>Results</jats:title> <jats:p> At lesion level, RG-SSL-LOC achieved higher median Dice than FSL and RG-SSL (0.49 vs 0.41 and 0.40; both <jats:italic>p</jats:italic> <.001). At case level, RG-SSL-LOC achieved area-under-the-curve (AUC) values of 0.83, 0.82, and 0.87 in the external retrospective, external prospective, and internal prospective cohorts, respectively. Compared with FSL, AUCs were 0.84 ( <jats:italic>p</jats:italic> =.237), 0.80 ( <jats:italic>p</jats:italic> =.020), and 0.84 ( <jats:italic>p</jats:italic> <.001); compared with RG-SSL, AUCs were 0.83 ( <jats:italic>p</jats:italic> =.929), 0.82 ( <jats:italic>p</jats:italic> =.652), and 0.86 ( <jats:italic>p</jats:italic> =.007); compared with PI-RADS, AUCs were 0.78 ( <jats:italic>p</jats:italic> =.055), 0.83 ( <jats:italic>p</jats:italic> =.652) and 0.86 ( <jats:italic>p</jats:italic> =.480). Combined with clinico-radiological variables, RG-SSL-LOC significantly improved AUC versus clinico-radiological variables alone in the external retrospective (0.85 vs 0.80, <jats:italic>p</jats:italic> =.002), external prospective (0.87 vs 0.84, <jats:italic>p</jats:italic> =.008), and internal prospective (0.91 vs 0.88, <jats:italic>p</jats:italic> <.001) cohorts; in the latter, it reduced unnecessary biopsies by 15.19%. </jats:p> </jats:sec> <jats:sec> <jats:title>Conclusion</jats:title> <jats:p>RG-SSL-LOC achieves better segmentation quality than other methods, demonstrates robust prospective multicenter performance and improves multimodal detection.</jats:p> </jats:sec> <jats:sec> <jats:title>Summary</jats:title> <jats:p>A report-guided semi-supervised method outperforms fully-supervised baseline on prospective multicenter data for prostate cancer detection and adds value in multimodal approaches, effectively using unlabelled data and facilitating model scaling.</jats:p> </jats:sec> <jats:sec> <jats:title>Key Points</jats:title> <jats:list list-type="bullet"> <jats:list-item> <jats:p>The use of a lesion-only teacher model in a state-of-the-art report-guided semi-supervised learning framework improved prostate cancer segmentation quality and detection performance in biparametric MRI</jats:p> </jats:list-item> <jats:list-item> <jats:p>Report-guided semi-supervised learning more than tripled the amount of training data available, with the proposed lesion-only teacher approach retaining 7.02% more malignant cases than the state-of-the-art approach.</jats:p> </jats:list-item> <jats:list-item> <jats:p>In an internal prospective cohort, the proposed multimodal approach could potentially reduce unnecessary biopsies by 15.2%.</jats:p> </jats:list-item> </jats:list> </jats:sec>