Prediction of Prostate Cancer Grades Using Radiomic Features

We developed a machine learning model for predicting prostate cancer (PCa) grades using radiomic features of magnetic resonance imaging. 112 patients diagnosed with PCa based on prostate biopsy between January 2014 and December 2021 were evaluated. Logistic regression was used to construct two prediction models, one using radiomic features and prostate-specific antigen (PSA) values (Radiomics model) and the other Prostate Imaging-Reporting and Data System (PI-RADS) scores and PSA values (PI-RADS model), to differentiate high-grade (Gleason score [GS] >= 8) from intermediate or low-grade (GS <8) PCa. Five imaging features were selected for the Radiomics model using the Gini coefficient. Model performance was evaluated using AUC, sensitivity, and specificity. The models were compared by leave-one-out cross-validation with Ridge regularization. Furthermore, the Radiomics model was evaluated using the holdout method and represented by a nomogram. The AUC of the Radiomics and PI-RADS models differed significantly (0.799, 95% CI: 0.712-0.869; and 0.710, 95% CI: 0.617-0.792, respectively). Using holdout method, the Radiomics model yielded AUC of 0.778 (95% CI: 0.552-0.925), sensitivity of 0.769, and specificity of 0.778. It outperformed the PI-RADS model and could be useful in predicting PCa grades, potentially aiding in determining appropriate treatment approaches in PCa patients.