Predicting Genetic Subtypes of Prostate Cancer Using Radiomics Features from T2-weighted Prostate MRI Images

Motivation: Prostate cancer has diverse genetic subtypes affecting prognosis and treatment response. Goal(s): Develop a machine learning model to predict four genetic subtypes (Luminal A, Luminal S, AVPC-I, ACPV-M) using radiomic features from T2-weighted MRI, supporting personalized treatment. Approach: In 195 patients, RNA sequencing identified subtypes. T2-weighted MRIs were segmented, and 1,422 radiomic features were extracted. Feature selection used ICC, and classification models were trained with SMOTE for data balance. Results: The RF model achieved AUROC scores of 0.98 (train) and 0.84 (test) for AR inhibitor-resistant subtype, and 0.95 (train) and 0.77 (test) for differentiating docetaxel-resistant subtypes. Impact: Our pilot study demonstrated promising results for a radiomics model capable of predicting genetic subtypes of prostate cancer. This model demonstrated its ability to predict AR inhibitor-resistant and docetaxel-resistant genetic subtypes with AUROCs of 0.84 and 0.77, respectively.