To explore the value of radiomics in stratifying the Gleason score (GS) of prostate cancer based on vast image features from biparametric MRI.

Three hundred and sixteen patients were enrolled in this study from October, 2015 to December, 2018 and their results of surgical pathology were obtained. The lesions were manually depicted by 3D-Slicer. Then, 106-dimensional features extracted by radiomics were used to conduct Spearman non-parametric correlation test with the high and low risk stratification of GS. The constructed Neural Network was trained with the features after dimension reduction by principal component analysis as the input. Then, the testing set was fed in to get the predictive capability of the model. In the end, 10-fold cross-validation and shuffle of 100 times were used to test the accuracy of the prediction and the generalization ability of the model.

Seventy seven-dimensional features with significant correlation were found at the level of P valued=0.05 (two-tailed). After dimensional features were reduced, 21 dimensional new feature spaces with 99% original feature information were obtained. The results on the testing data after the 10-fold validation and shuffle were AUC=0.712 with T₂WI, AUC=0.689 with DWI (b=1 000 s/mm²), AUC=0.689 with DWI (b=2 000 s/mm²) and AUC=0.691 with DWI (b=3 000 s/mm²).

The neural network after extracting features from biparametric MRI images can accurately and automatically distinguish the high risk and low risk groups of Gleason grade of prostatic cancer.

Prostatic neoplasms; Magnetic resonance imaging; Radiomics