Uncertainty-aware deep learning for classifying colonic fdg-pet uptake using radiomic features.

This study developed deep learning models to differentiate tumorous from non-tumorous incidental colorectal FDG-PET uptake using radiomic features categories and to evaluate both classification performance and predictive uncertainty. PET data from 80 patients (45 tumorous, 35 non-tumorous) were analyzed. Radiomic features were grouped into morphology, intensity, texture, and all features, and feature selection was performed using LassoCV within each training split. Model development and evaluation used repeated stratified hold-out validation (10 splits; train/test = 80/20) with internal cross-validation and early stopping. Predictive uncertainty was estimated using Monte Carlo dropout with 100 stochastic inferences per test case and assessed using risk–coverage curves and AURC. The All model achieved the highest accuracy (84.8 ± 5.7%) and AUC (0.83 ± 0.08), outperforming single-category models (accuracy < 70%; AUC 0.70–0.74). Uncertainty increased as predicted probabilities approached 0.5, and selective prediction based on uncertainty reduced misclassification risk for the Intensity, Texture, and All models, but not for the Morphology model. These results suggest that integrating multiple radiomic feature categories improves both predictive performance and stability, and that uncertainty-guided decision-making may support safer clinical triage by withholding high-uncertainty cases.