Chapter 14 - Deep learning contributions for reducing the complexity of prostate biomechanical models

Abstract

Prostate cancer (PCa) is the most common malignancy in males in western countries, which is usually detected with prostate-specific antigen (PSA) levels measurement, along with digital rectal examination, and Ultrasound (US) systematic biopsy for confirmation. Recently, multi-parametric Magnetic Resonance Imaging (mpMRI) has improved PCa diagnosis significantly, allowing for accurate, noninvasive detection of PCa lesions, and opening the door to MR-guided US biopsies that can directly target the lesions, as opposed to classical systematic biopsies which are expensive and error-prone. Targeted biopsies are usually done using very accessible transrectal US (TRUS) probes; it requires finding the correspondence between the pre-acquired prostate mpMRI and the intraoperative TRUS image, a problem known as MR-TRUS registration or fusion. In this chapter, an automatic system for near-real-time MR-TRUS prostate registration will be developed and validated using prostate MR-TRUS pairs from 204 patients. The dense deformation field (DDF) transforming a patient's MR prostate points (along with marked lesions) to the corresponding TRUS prostate points are calculated using Coherent Point Drift (CPD) to match prostate surfaces, followed by a Finite Element Method (FEM) simulation to obtain mechanically plausible internal deformations. Then, a Convolutional Neural Network (CNN) is trained to directly predict the DDFs from MR-US pairs (along with the corresponding prostate masks, which are automatically segmented), attaining an almost perfect approximation to the CPD+FEM DDF while reaching near-real-time speeds.