Male pelvic multi-organ segmentation using V-transformer network

Abstract

Automatic multi-organ segmentation is a cost-effective tool for generating organ contours using computed tomography (CT) images. This work proposes a deep-learning algorithm for multi-organ (bladder, prostate, rectum, left and right femoral heads) segmentation in pelvic CT images for prostate radiation treatment planning. We propose an encoder-decoder network with a V-net backbone for local feature extraction and contour reconstruction. Novel to our network, we utilize a token-based transformer, which encourages long-range dependency, to forward more informative high-resolution feature maps from the encoder to the decoder. In addition, a knowledge distillation strategy was applied to improve the network’s generalization. We evaluate the network using a dataset collected from 50 patients with prostate cancer. A quantitative evaluation of the proposed network’s performance was performed on each organ based on: 1) volume similarity between the segmented contours and ground truth using Dice score, segmentation sensitivity, precision, and absolute percentage volume difference (AVD), 2) surface similarity evaluated by Hausdorff distance (HD), mean surface distance (MSD) and residual mean square distance (RMSD). The performance was then evaluated against other state-of-art methods. The average volume similarities achieved by the network over all organs were: Dice score = 0.83, sensitivity = 0.84, and precision = 0.83; the average surface similarities were HD = 5.77mm, MSD = 0.93mm, RMSD = 2.77mm, and AVD =12.85%. The proposed methods performed significantly better than competing methods in most evaluation metrics. The proposed network may be a promising segmentation approach for use in routine prostate radiation treatment planning.