To develop and internally validate a PSMA PET/CT–centric deep-learning framework based on using a foundational model for noninvasive preoperative risk stratification of prostate cancer at the ISUP Grade Group (GG) 1–2 vs. 3–5 boundary, without using biopsy-derived inputs at inference. In this single-center retrospective cohort, 494 men underwent 18 FPSMA-1007 PET/CT within one month before radical prostatectomy. Intraprostatic ROIs were manually delineated. We developed a dual-path hybrid model that fuses global semantic features from a frozen BiomedCLIP foundation model with task-specific 3D PET/CT features, and used the fused representation for GG 1–2 vs. 3–5 stratification. Five-fold, patient-level cross-validation was used; prostatectomy pathology served as the reference standard. Benchmarks included conventional radiomics and deep-learning baselines (BiomedCLIP-only, ResNet-only, MedSAM, XSurv). The primary metric was AUC; precision, recall, F1, PRC-AUC, and decision-curve analysis (DCA) assessed complementary performance and clinical utility. The proposed model achieved AUC 0.800 with precision 0.854, recall 0.888, and F1 0.870, outperforming baselines (BiomedCLIP 0.764; MedSAM 0.759; XSurv 0.756; ResNet 0.745; radiomics/random forest 0.676). Mean PRC-AUC was 0.928 ± 0.020 (across folds), and DCA showed higher net benefit across wide thresholds. SUVmax alone was modest (AUC 0.699 for ≥GG3; 0.625 for ≥GG4). The framework demonstrated noninvasive discrimination of GG 1–2 vs. 3–5 in a single-center cohort, suggesting a candidate decision-support role for biopsy-sparing pathways; its generalizability and deployability require multicenter external validation and end-to-end automation.
Bian et al. (Mon,) studied this question.