Prostate Cancer Detection and Segmentation in Multi-parametric MRI via CNN and Conditional Random Field

Abstract

Multi-parametric MRI (mp-MRI) is a powerful diagnostic tool for prostate cancer (PCa). However, interpreting prostate mp-MRI requires high-level expertise, causing significant inter-reader variations. Convolutional neural networks (CNNs) have recently shown great promise for various tasks. In this study, we propose an improved CNN to jointly detect PCa lesions and segment for accurate lesions contours. Specifically, we adapt focal loss to overcome the imbalance between cancerous and non-cancerous areas for improved lesion detection and design selective dense conditional random field (SD-CRF), a post-processing step to refine the CNN prediction into the lesion segmentation based on a specific imaging component of mp-MRI. We trained and validated the proposed CNN in 5-fold cross-validation using 397 preoperative mp-MRI exams with whole-mount histopathology-confirmed lesion annotations. In the free-response receiver operating characteristics (FROC) analysis, the proposed CNN achieved 75.1% lesion detection sensitivity at the cost of 1 false positive per patient. In the evaluation for lesion segmentation, the proposed CNN improved the Dice coefficient by 20.6% from the baseline CNN.