Automated PI-RADS 3–5 Classification Using Multiparametric MRI: A Comparative Study of Radiomics and Deep Learning Approaches

Accurate classification of clinically significant prostate cancer remains a major challenge. While multiparametric MRI (mpMRI) has improved lesion detection, effective categorization in accordance to the Prostate Imaging Reporting and Data System (PI-RADS) remains complex. In this study, we propose and evaluate three complementary approaches for automated PI-RADS classification differing in the way in which the features are extracted from the mpMRI imaging sequences. The first approach extracts hand-crafted radiomic features from manually segmented lesions using the PyRadiomics library. The second approach extends this by integrating fully automated lesion and zonal segmentation to simulate a real-world, manual-free pipeline. The third approach utilizes a custom convolutional neural network (CNN) to learn high-level features images and lesion masks directly. The images come from Apparent diffusion coefficient (ADC), diffusion-weighted imaging (DWI), and T2-weighted (T2W) imaging. The features issued by the three methods were used to train a set of machine learning models for multi-class PI-RADS classification, specifically targeting the clinically relevant categories 3, 4, and 5. Results show that ADC-derived features consistently yield superior performance, with one of the ensemble models reaching an AUC of 0.83. Combining features across all sequences further improved robustness (AUC = 0.84). PI-RADS 5 classification was most reliable (AUC ≥ 0.94), whereas PI-RADS 3 remained the most difficult to distinguish. Our findings highlight the effectiveness of ADC features and the advantage of combining automated and deep learning-based strategies for robust prostate cancer risk stratification.