Dual Shape Guided Segmentation Network for Organs-at-Risk in Head and Neck CT Images

TL;DR

This paper introduces DSGnet, a dual shape guided segmentation network that improves automatic delineation of organs-at-risk in head and neck CT images by leveraging shape representations and shared features, achieving high accuracy.

Contribution

The proposed DSGnet uniquely combines shape guidance via UIDM with shared feature learning to handle shape variation and boundary ambiguity in OAR segmentation.

Findings

Achieved an overall DSC of 0.842 across nine OARs.

Outperformed state-of-the-art methods in accuracy and efficiency.

Built a large dataset of 699 head and neck CT images.

Abstract

The accurate segmentation of organs-at-risk (OARs) in head and neck CT images is a critical step for radiation therapy of head and neck cancer patients. However, manual delineation for numerous OARs is time-consuming and laborious, even for expert oncologists. Moreover, manual delineation results are susceptible to high intra- and inter-variability. To this end, we propose a novel dual shape guided network (DSGnet) to automatically delineate nine important OARs in head and neck CT images. To deal with the large shape variation and unclear boundary of OARs in CT images, we represent the organ shape using an organ-specific unilateral inverse-distance map (UIDM) and guide the segmentation task from two different perspectives: direct shape guidance by following the segmentation prediction and across shape guidance by sharing the segmentation feature. In the direct shape guidance, the segmentation prediction is not only supervised by the true label mask, but also by the true UIDM, which is implemented through a simple yet effective encoder-decoder mapping from the label space to the distance space. In the across shape guidance, UIDM is used to facilitate the segmentation by optimizing the shared feature maps. For the experiments, we build a large head and neck CT dataset with a total of 699 images from different volunteers, and conduct comprehensive experiments and comparisons with other state-of-the-art methods to justify the effectiveness and efficiency of our proposed method. The overall Dice Similarity Coefficient (DSC) value of 0.842 across the nine important OARs demonstrates great potential applications in improving the delineation quality and reducing the time cost.