Matching axial images of magnetic resonance imaging and transrectal ultrasound based on deep learning

Abstract

This paper examines the feasibility of automated alignment in prostate targeted biopsy by comparing the prostate contour between different modalities. The prostate targeted biopsy that is attracting attention in the treatment of prostate cancer largely depends on the doctor who operates surgery, so it can be expected to reduce the variation in the diagnostic performance by automation. In the proposed method, segmentation is performed using deep learning, and the same prostate cross section between different modalities is estimated from the similarity obtained by comparing prostate contours of different modalities obtained by segmentation. In this method it was possible to estimate close to expert judgment with accuracy of 69.4%. Furthermore, by considering the deformity of the prostate gland and calculating the similarity for each angle, we achieved an estimate close to the judgment of experts with higher accuracy of 83.3%.