Poster Session 7: Oncology – Prostate (Part 2)

Introduction:We aimed to develop a radiomics-based prognostic model using machine learning to predict lymph node invasion (LNI), biochemical recurrence (BCR), metastasis-free survival (MFS), definitive androgen deprivation therapy (dADT)-free survival (FS), castration-resistant prostate cancer (CRPC)-FS, and prostate cancer-specific survival (PCSS) in individuals diagnosed with high-grade prostate cancer (PCa).Methods: A total of 295 individuals with high-grade PCa (Gleason score ≥8) underwent preoperative positron emission tomography (PET) with 18F-fluorodeoxyglucose (FDG) combined with computed tomography (CT) imaging at our tertiary care health center in Quebec City, Canada.Clinical data (CD), including age, prostate-specific antigen (PSA) level, Gleason score, and clinical stage, were used to build prognostic models to which handcrafted radiomics (HCR) or deep learning-based radiomics (DLR) were added to enhance performance.We trained the models using a subset of the cohort (250 individuals) and optimized them using a stratified five-fold cross-validation.The selected model was then validated using a test set of 45 individuals.Performance on the test set was evaluated using the area under the curve of the receiver operator characteristic (AUC-ROC) and the concordance index (C-index).A comparison with commonly used nomograms (MSKCC and CAPRA-S) was also made.Results: Median followup was 64.7 (range 29.3-89.6)months.Median age was 66 (48-80) years. .A total of 230 (88%) and 31 (12%) had clinical T1-T2 and T3a disease, respectively.The majority (63.7%) had Gleason 8.At RP, 86 (29%) individuals had LNI.At followup, 160 had BCR.In the training set, using CD with radiomics yielded better performance for prediction of LNI (AUC 72±5 vs. 70±7 MSKCC and 62±4 CAPRA-S) and BCR-FS (CI=65±6 vs. 64±3 MSKCC and 63±4 CAPRA-S).Nomograms outperformed our combined radiomics-CD model for prediction of other outcomes, although performances were like our CD-only model.Conclusions: Integrating imaging data into prognostic tools through artificial intelligence enhances clinical predictions for LNI and BCR, enabling more accurate prognostication.With minimal training, we achieved better results than commonly used nomograms.