Purpose: To predict Gleason Score (GS) using radiomic features from ⁶⁸Ga-PSMA-PET/CT images in primary prostate cancer.

Methods: 138 patients undergoing ⁶⁸Ga-PSMA-PET/CT imaging were categorized based on GS, with GS above 4 + 3 as malignant and under 3 + 4 as benign tumors. radiomic features were extracted from tumors’ volume of interest in both PET and CT images, using Feature Elimination with cross-validation. Fusion features were generated by combining features at the feature level; average of features (PET/CT_AveFea) or concatenated features (PET/CT_ConFea). The performance of various models was compared using area under the curve, sensitivity and specificity. Wilcoxon test and F1-score test were used to find the best model. Predictive models were developed for CT-only, PET-only, and PET/CT feature-level fusion models.

Results: Random Forest achieved the highest accuracy on CT with 0.74 ± 0.01 AUC_Mean, 0.75 ± 0.07 sensitivity, and 0.62 ± 0.08 specificity. Logistic regression (LR) exhibited the best predictive performance on PET images with 0.74 ± 0.05 AUC_Mean, 0.7 ± 0.13 sensitivity, and 0.78 ± 0.14 specificity. The best predictive PET/CT_AveFea was achieved by LR, resulting in 0.72 ± 0.07 AUC_Mean, 0.74 ± 0.12 sensitivity, and 0.63 ± 0.02 specificity. In the case of PET/CT_ConFea, LR showed the best predictive performance with 0.78 ± 0.08 AUC_Mean, 0.81 ± 0.09 sensitivity, and 0.66 ± 0.15 specificity.

Conclusion: The results demonstrated that radiomic models derived from ⁶⁸Ga-PSMA-PET/CT images could differentiate between benign and malignant tumors based on GS.