Dual-branch feature Reinforcement Transformer for preoperative parathyroid gland segmentation

Lei Lyu, Chen Pang, Qinghan Yang, Kailin Liu, Chong Geng

Abstract: The parathyroid glands are easily injured or accidentally removed during thyroid surgery, causing temporary or even permanent hypocalcemia. Thus, accurate preoperative identification and localization of the parathyroid glands by ultrasound is crucial in protecting the parathyroid glands and preventing parathyroid injury during thyroid surgery. However, there are only a few methods used for highlighting the parathyroid gland in ultrasound images before thyroid surgery. In this study, we propose a Dual-branch feature Reinforcement Transformer Network (DRT-Net) for preoperative parathyroid gland segmentation. DRT-Net incorporates a dual-branch structure, consisting of a devised convolution network (CNN) backbone called Feature Reinforcement subnet (FR-subnet) and a Transformer branch capturing detailed features and context information from the confused ultrasound image. Furthermore, we design a Margin Tracking Attention (MTA) that optimizes the ability of FR-subnet to process margin information by tracking margin pixels of feature map. Finally, we employ a Cross-channel Feature Reinforcement Module (CFRM) to fuse the extracted detailed features from the CNN branch with the global context information from the Transformer branch. We trained and evaluated the DRT-Net on the self-built parathyroid gland segmentation dataset and an open-access Kvasir-SEG dataset. Extensive experiments have been carried out to validate the efficiency of our method.