BDU‐net: Toward accurate segmentation of dental image using border guidance and feature map distortion

Abstract

AbstractThe importance of instance segmentation for each tooth is increasing in dental disease diagnosis and computer‐assisted treatment. However, most existing segmentation methods are mainly concerned with semantic feature extraction, while ignoring complex situations such as blurred boundaries and dislocation of teeth in panoramic radiographs. To address these problems, we propose a dual subnetworks structure based on border guidance and feature map distortion, called BDU‐net. Specifically, we embed the Disout method with the capability to distort features into the encoder to obtain multi‐scale feature maps with excellent generalization. Then, the distorted feature map is submitted to two subnetworks, each having a feature pyramid structure in a shared manner, including a border subnetwork that adjusts the segmentation boundaries and a region subnetwork that generates the region segmentation results. Experiments conducted on the teeth segmentation dataset show that the proposed BDU‐net performs better than the other approaches, especially for the case of teeth missing or dislocated teeth.