Differential Diagnosis of Prostate Volume Images by Transfer Learning.

It is difficult to differentiate Prostate Cancer from Prostate Hyperplasia in medical images. We assess the utility of deep learning techniques applied to Magnetic Resonance Diffusion images to improve diagnosis. We initially deploy a 2.5D volume-image technique for ResNet-50 retraining and subsequently a 2D+1D volume fly-through for MobileNet-3D retraining. Both techniques result in an 80% Accuracy, which is comparable to the reported Radiologist's Accuracy of 79%. We have not improved the differential diagnosis, but the techniques show potential and may improve as more quality-assured images become available.