Prostate cancer detection and segmentation on MRI using non-local mask R-CNN with histopathological ground truth

Abstract

Background: Automatic detection and segmentation of intraprostatic lesions (ILs) on preoperative multiparametric-magnetic resonance images (mp-MRI) can improve clinical workflow efficiency and enhance the diagnostic accuracy of prostate cancer and is an essential step in dominant intraprostatic lesion boost.

Purpose: The goal is to improve the detection and segmentation accuracy of 3D ILs in MRI by a proposed a deep learning (DL)-based algorithm with histopathological ground truth.

Methods: This retrospective study included 262 patients with in vivo prostate biparametric MRI (bp-MRI) scans and were divided into three cohorts based on their data analysis and annotation. Histopathological ground truth was established by using histopathology images as delineation reference standard on cohort 1, which consisted of 64 patients and was randomly split into 20 training, 12 validation, and 32 testing patients. Cohort 2 consisted of 158 patients with bp-MRI based lesion delineation, and was randomly split into 104 training, 15 validation, and 39 testing patients. Cohort 3 consisted of 40 unannotated patients, used in semi-supervised learning. We proposed a non-local Mask R-CNN and boosted its performance by applying different training techniques. The performance of non-local Mask R-CNN was compared with baseline Mask R-CNN, 3D U-Net and an experienced radiologist's delineation and was evaluated by detection rate, dice similarity coefficient (DSC), sensitivity, and Hausdorff Distance (HD).

Results: The independent testing set consists of 32 patients with histopathological ground truth. With the training technique maximizing detection rate, the non-local Mask R-CNN achieved 80.5% and 94.7% detection rate; 0.548 and 0.604 DSC; 5.72 and 6.36 95 HD (mm); 0.613 and 0.580 sensitivity for ILs of all Gleason Grade groups (GGGs) and clinically significant ILs (GGG > 2), which outperformed baseline Mask R-CNN and 3D U-Net. For clinically significant ILs, the model segmentation accuracy was significantly higher than that of the experienced radiologist involved in the study, who achieved 0.512 DSC (p = 0.04), 8.21 (p = 0.041) 95 HD (mm), and 0.398 (p = 0.001) sensitivity.

Conclusion: The proposed DL model achieved state-of-art performance and has the potential to help improve radiotherapy treatment planning and noninvasive prostate cancer diagnosis.

Keywords: deep learning; detection and segmentation; prostate cancer.