Contributions of biomechanical modeling and machine learning to the automatic registration of multiparametric magnetic resonance and transrectal echography for prostate brachytherapy

Abstract

Prostate cancer (PCa) is the most common malignancy in western males, and third by mortality. After detecting elevated Prostate Specific Antigen (PSA) blood levels or after a suspicious rectal examination, a Magnetic Resonance (MR) image of the prostate is acquired and assessed by radiologists to locate suspicious regions. These are then biopsied, i.e. living tissue samples are collected and analyzed histopathologically to confirm the presence of cancer and establish its degree of aggressiveness. During the biopsy procedure, Ultrasound (US) is typically used for guidance and lesion localization. However, lesions are not directly visible in US, and the urologist needs to use fusion software to performs MR-US registration, so that the MR-marked locations can be transferred to the US image. This is essential to ensure that the collected samples truly come from the suspicious area. This work compiles five publications employing several Artificial Intelligence (AI) algorithms to analyze prostate images (MR and US) and thereby improve the efficiency and accuracy in diagnosis, biopsy and treatment of PCa: 1. Automatic prostate segmentation in MR and US: Prostate segmentation consists in delimiting or marking the prostate in a medical image, separating it from the rest of the organs or structures. Automating this task fully, which is required for any subsequent analysis, saves significant time for radiologists and urologists, while also improving accuracy and repeatability. 2. Segmentation resolution enhancement: A methodology for improving the resolution of the previously obtained segmentations is presented. 3. Automatic detection and classification of MR lesions: An AI model is trained to detect lesions as a radiologist would and to estimate their risk. The model achieves improved diagnostic accuracy, resulting in a fully automatic system that could be used as a second clinical opinion or as a criterion for patient prioritization. 4. Simulation of biomechanical behavior in real time: It is proposed to accelerate the simulation of biomechanical behavior of soft organs using AI. 5. Automatic MR-US registration: Registration allows localization of MR-marked lesions on US. High accuracy in this task is essential for the correctness of the biopsy and/or focal treatment procedures (such as high-rate brachytherapy). Here, AI is used to solve the registration problem in near-real time, while exploiting underlying biomechanically-compatible models.