Prostate MRI and PI-RADS play a central role in identifying clinically significant prostate cancer (sPCa), but false-positive findings can lead to unnecessary biopsies. This study explored whether quantitative MRI radiomics and clinical information can complement PI-RADS to improve risk prediction.

Using 1,497 prostate MRI examinations from 1,395 men, the researchers developed and validated four machine learning models based on PI-RADS, radiomics, their combination, and a multimodal approach integrating PI-RADS, radiomic features and clinical variables. Radiomic features were extracted from automatically segmented prostate regions on T2-weighted, DWI and ADC images.

While radiomics alone performed similarly to PI-RADS, combining imaging and clinical information delivered the strongest results. The PI-RADS + radiomics + clinical model achieved an AUC of 0.891, significantly outperforming the other approaches.

At a fixed sensitivity of 97.4%, this integrated model also achieved the highest specificity (29.4%) and biopsy avoidance rate (18.15%), while maintaining the same rate of undetected sPCa.

The findings highlight the value of multimodal AI for prostate cancer diagnosis, showing how quantitative MRI biomarkers can be combined with established radiological assessment and clinical data to refine sPCa risk stratification and potentially reduce unnecessary prostate biopsies.