Synolitic Graph Neural Networks for MRI-Derived Radiomic-Based Prediction of Prostate Cancer Progression on Active Surveillance

Background: Prostate cancer (PCa) is one of the most prevalent malignancies in men, and active surveillance (AS) is the recommended management strategy for low- and favourable intermediate-risk disease. Predicting which patients will progress during AS remains a clinical challenge. MRI-derived radiomics has shown promise for risk stratification, but conventional machine learning approaches treat radiomic features as independent variables and may not capture the complex inter-feature dependencies within imaging data. This study evaluates the application of Synolitic Graph Neural Networks (SGNNs) to baseline MRI-derived radiomic features for predicting prostate cancer progression on active surveillance. Methods: We studied 343 AS patients (73 progressors, 270 non-progressors) from a single-centre cohort prospectively enrolled between 2013 and 2019 and retrospectively analysed. Seventy-two radiomic features were extracted from baseline 3T MRI (T2-weighted imaging and apparent diffusion coefficient maps), together with three clinical variables (prostate volume, PSA, PSA density). The SGNN pipeline transformed each patient’s feature profile into a weighted graph encoding pairwise feature relationships via logistic regression classifiers trained within each cross-validation fold. GCN and GATv2 architectures were evaluated with multiple sparsification strategies and compared against Gradient Boosting, SVM, Random Forest, and logistic regression using 5-fold stratified cross-validation. Results: Among conventional methods, Gradient Boosting achieved the highest ROC-AUC (0.634 ± 0.080). The SGNN pipeline with GATv2, confidence-based sparsification (p = 0.8), and extended node features incorporating graph centrality measures achieved the best performance (ROC-AUC = 0.699 ± 0.044), an absolute improvement of 0.065 over the best conventional method. The addition of topological node features consistently improved performance by 3–5% across configurations. GATv2 outperformed GCN in matched settings. Conclusions: As a proof of concept, the SGNN framework achieved the highest mean ROC-AUC among the evaluated single-timepoint approaches, though results require confirmation in independent external cohorts. By encoding inter-feature relationships as patient-specific graphs, SGNN offers a complementary modelling paradigm for radiomic data in clinical oncology. Future work should incorporate longitudinal data and graph explainability methods.