Differentiating Prostate Cancer from Normal Tissue Using Radiomics from Ga-68 PSMA PET/CT: A Comparative Study of Twelve Machine Learning Classifiers

Abstract To evaluate the feasibility and robustness of radiomics-based machine learning on gallium-68 prostate-specific membrane antigen positron emission tomography/computed tomography for differentiating histopathologically confirmed prostate cancer from normal prostatic tissue, with emphasis on control-group composition and class balancing. This retrospective study included 73 subjects (38 prostate cancer, 35 controls). Manual prostate segmentation yielded 186 radiomics features, of which 127 remained after removal of invalid and zero-variance features. Four analyses were performed: 38 cancer/14 normal with synthetic minority oversampling technique (SMOTE), 38/24 with SMOTE, 38/35 with SMOTE, and 38/35 without SMOTE. Twelve classifiers were evaluated using 10,000 repeated random 80/20 train-test partitions. Balanced accuracy was the primary metric. Classifier differences were assessed using paired bootstrap confidence intervals (CIs) and sign-flipping permutation tests with Benjamini–Hochberg correction. Performance varied with control-group composition. CoarseKNN achieved the highest balanced accuracy in the 38/14 analysis (0.880 ± 0.111) and both 38/35 analyses (0.817 ± 0.097 with SMOTE; 0.812 ± 0.097 without SMOTE). In the 38/24 analysis, MediumKNN performed best (0.810 ± 0.113), followed by CosineKNN (0.807 ± 0.113) and CoarseKNN (0.798 ± 0.115). MediumKNN significantly outperformed CoarseKNN (mean paired difference, 0.0126; 95% CI, 0.0108–0.0144; false discovery rate-adjusted p < 0.001). K-nearest neighbors (KNN) classifiers showed the most consistent performance, although the best classifier varied with cohort composition. The biopsy-negative analysis provided a clinically relevant control comparison. However, the small sample size, heterogeneous nonbiopsy controls, and absence of external validation limit generalizability. Larger prospective multicenter studies with direct benchmarking against PROMISE and PRIMARY score are required before clinical translation.