Deep learning and Magnet Resonance Imaging for Prostate Cancer Detection and Determination of the clinical Significance

Abstract Background A human perception-based assessment of multi-parametric magnetic resonance imaging (mpMRI) of the prostate does not necessarily tap the full potential in determining prostate cancer (PCa) and identifying significant prostate cancer (sPCa). Methods Our multi-institutional international study includes 6,448 mpMRI prostate images from 1,830 patients (PCa diagnosis in 69.7% of patients). MR Images from a single institution were utilized for the model development and in-house validation, and from two international institutions for external validation. We utilized volumetric data, PlexusNET architecture, and attention algorithms to develop deep learning models. Performance was measured using the area under receiving characteristic operating curve (AUROC) and compared to the PI-RADS score system (version 2) at the case level for PCa diagnosis and sPCa identification. The reduction rate of biopsy settings without missing any PCa cases measured the clinical utility. Results Our compact models were internally and externally validated for a significant improvement in PCa detection by 7.25% compared to the PI-RADS score system. Following the model recommendation would avoid at least 11.3% of unnecessary biopsies. Moreover, the DL model correctly predicted PCa presence in 22.5% of cases, which were misclassified according to the PI-RADS score system. The identification accuracy of sPCa for the model was statistically significantly higher than PI-RADS scores (AUROC: 0.769 vs. 0.726; p < 0.021) on a PCa cohort with 79% sPCa. Conclusions Our solution facilitates mpMRI assessment of the prostate for PCa diagnosis and the determination of sPCa; we demonstrated a great potential of AI for clinical utility and improved mpMRI assessment.