Purpose: To explore the potential of automatic stratification of prostate cancer patients into low- and high-grade groups using a support vector machine (SVM) model with multiparametric magnetic resonance (MR) image features.
Methods: MR images [T1- and T2-weighted, diffusion weighted b=0, b=1,000, b=2,000, and apparent diffusion coefficient (ADC) images] of 101 cancer regions corresponding to histopathological images of prostate cancer were selected and divided into a training (n=72) and test datasets (n=29). Each dataset was divided into two groups, low-grade tumors [≤ International Society of Urological Pathology (ISUP) Grade Group (GG) 2] and high-grade tumors (ISUP GG3≤). 1,620 MR image features were derived from original images and two-dimensional wavelet filtered images for the six types of images. Seven significant features were chosen by using a least absolute shrinkage and selection operator (LASSO) algorithm to build a SVM model. Parameters of the SVM model were optimized based on a leave-one-out cross validation test so that areas under receiver operating characteristic curves (AUCs) can be maximized.
Results: The most significant feature was one of wavelet features from ADC images. The AUCs based on the SVM model of the automatic stratification of prostate cancer patients into two groups were 0.977 in the training dataset and 0.910 in the test dataset.
Conclusion: This study suggested the potential of automatic stratification of prostate cancer patients into low- and high-grade groups using the SVM model with multiparametric MR image features.