Detection of prostate cancer and determination of its significance using explainable artificial intelligence.

5555 Background: The variation of the human perception has limited the potential of multi-parametric magnetic resonance imaging (mpMRI) of the prostate in determining prostate cancer and identifying significant prostate cancer. The current study aims to overcome this limitation and utilizes an explainable artificial intelligence to leverage the diagnostic potential of mpMRI in detecting prostate cancer (PCa) and determining its significance. Methods: A total of 6,020 MR images from 1,498 cases were considered (1,785 T2 images, 2,719 DWI images, and 1,516 ADC maps). The treatment determined the significance of PCa. Cases who received radical prostatectomy were considered significant, whereas cases with active surveillance and followed for at least two years were considered insignificant. The negative biopsy cases have either a single biopsy setting or multiple biopsy settings with the PCa exclusion. The images were randomly divided into development (80%) and test sets (20%) after stratifying according to the case in each image type. The development set was then divided into a training set (90%) and a validation set (10%). We developed deep learning models for PCa detection and the determination of significant PCa based on the PlexusNet architecture that supports explainable deep learning and volumetric input data. The input data for PCa detection was T2-weighted images, whereas the input data for determining significant PCa include all images types. The performance of PCa detection and determination of significant PCa was measured using the area under receiving characteristic operating curve (AUROC) and compared to the maximum PiRAD score (version 2) at the case level. The 10,000 times bootstrapping resampling was applied to measure the 95% confidence interval (CI) of AUROC. Results: The AUROC for the PCa detection was 0.833 (95% CI: 0.788-0.879) compared to the PiRAD score with 0.75 (0.718-0.764). The DL models to detect significant PCa using the ADC map or DWI images achieved the highest AUROC [ADC: 0.945 (95% CI: 0.913-0.982; DWI: 0.912 (95% CI: 0.871-0.954)] compared to a DL model using T2 weighted (0.850; 95% CI: 0.791-0.908) or PiRAD scores (0.604; 95% CI: 0.544-0.663). Finally, the attention map of PlexusNet from mpMRI with PCa correctly showed areas that contain PCa after matching with corresponding prostatectomy slice. Conclusions: We found that explainable deep learning is feasible on mpMRI and achieves high accuracy in determining cases with PCa and identifying cases with significant PCa.