A Deep Learning Approach for Predicting Clinically Significant Prostate Cancer: A Retrospective, Multicentre Study

Purpose: To construct deep learning (DL) models based on multicentre biparametric MRI (bpMRI) for the diagnosis of clinically significant prostate cancer (csPCa), and compare the performance of these models with that of Prostate Imaging Reporting and Data System (PI-RADS) assessment of expert-level radiologists based on multiparametric MRI (mpMRI).Methods: This study included 1861 consecutive men with mpMRI from seven hospitals, who underwent radical prostatectomy or biopsy. These patients were divided into training cohort (3 hospitals, 1216 patients) and external validation cohorts (4 hospitals, 645 patients). PI-RADS assessment was performed by expert-level radiologists. The DL models were constructed for the classifications between benign and malignant lesions (DL-BM), and between csPCa and non-csPCa (DL-CS). The integrated model combining a DL-CS model and PI-RADS (PIDL-CS) was constructed. The performance of the deep learning models and PIDL-CS were compared with those of PI-RADS.Results: In each external validation cohort, the area under receiver operating characteristic curve (AUC) values of DL-BM and DL-CS models were not significantly different from that of PI-RADS (Ps > 0.05), whereas the AUC of PIDL-CS was superior to that of PI-RADS (Ps < 0.05) except one external validation cohort (P > 0.05). The specificity of the PIDL-CS for the detection of csPCa was much more than that of PI-RADS (Ps < 0.05).Conclusion: Our proposed DL model can be a potential non-invasive auxiliary tool to predict csPCa. Further, PIDL-CS greatly increased the specificity in the detection of csPCa compared with PI-RADS assessment of expert-level radiologists, greatly reducing the unnecessary biopsies and helping radiologists achieve a precise diagnosis of csPCa.