Introduction: To develop an automated multi-task prognostic model that combines clinical data with radiomics from positron emission tomography (PET) with 18F-fluorodeoxyglucose (FDG) combined with computed tomography (CT), eliminating the need for manual segmentation while providing clinically interpretable results. This is the first study of its kind using radiomics in prostate cancer that describes long-term clinical outcomes.  

Methods: 295 individuals with high-grade prostate cancer (Gleason score ≥8) who underwent radical prostatectomy (RP) and FDG-PET/CT imaging preoperatively at our tertiary care health center. Clinical data (CD), including age, prostate-specific antigen (PSA) level, clinical stage, and Gleason grade, were collected. Six prognostic tasks were defined, including lymph node invasion (LNI), biochemical recurrence (BCR)-free survival (FS), metastasis-free survival (MFS), definitive androgen deprivation therapy (dADT)-FS, castration-resistant prostate cancer (CRPC)-FS, and prostate cancer-specific survival (PCSS). A Bayesian Sequential Network (BSN), a dynamic prediction model quantifying uncertainty and adapting over time as outcomes from prior tasks unfold, was developed. It was compared with commonly used nomograms (MSKCC and CAPRA-S). Performance metrics on the holdout set were evaluated using the area under the curve of the receiver operator characteristic (AUC-ROC) and the concordance index (C-index).

Results: Median follow-up was 64.7 (range 29.3-89.6) months. Median age was 66 (48-80) years. Median PSA was 7.4 (1.1-155.3). 230 (88%) and 31 (12%) had clinical T1-T2 and T3a disease, respectively. At RP, 86 (29%) had LNI. At follow-up, 160 had BCR, 38 had metastases, 72 started dADT, 23 had CRPC, and 11 had PCSS. On the holdout set comprising 45 individuals, the BSN model outperformed nomograms for predicting LNI (AUC=66.3%), MFS (CI=75.3%), and dADT-FS (CI=69.6%). The nomogram still outperformed our BSN model for predicting BCR-FS (CI=63.5% [MSKCC] vs 59.2%), CRPC-FS (CI=67.6% [CAPRA-S] vs 65.6%), and PCSS (CI=87.8% [MSKCC] vs 78.0%).

Conclusion: Our fully automated self-learning multi-task model achieved good predictions with minimal training compared to commonly used nomograms while quantifying associated uncertainty.