SaTaNet: Structure-aware and Texture-aware end-to-end Network for cephalometric landmark localization on a new large dataset

A standard orthodontic diagnosis always requires cephalometric landmark localization to classify and quantify the anatomical abnormalities. The cephalometric X-ray image is a type of image with standard structure: specific fiducial point appears on specific area, and the topology of landmarks in different cephalometric X-ray images is basically similar. Existing methods primarily rely on multi-stage frameworks to progressively localize landmarks from coarse to fine scales; however, they fail to effectively integrate global topological cues with local texture information in an adaptive manner. In this work, we propose a structure-aware and texture-aware network, named SaTaNet, that can localize cephalometric landmark points automatically in a single-stage end-to-end network. The SaTaNet is guided by the topology information and texture information in the cephalometric X-ray to localize landmarks accurately. In order to make the network pay consistent attention of topology information and texture information across different scales, the multiscale attention module is proposed and added to SaTaNet. To evaluate our method, we construct a large-scale in-house dataset that covers a broad range of ages. The samples in the dataset are collected from 6128 patients. The patients were between 8 and 67 years of age. The annotations of the dataset passed the consistency test. Extensive experiments confirm that the proposed method delivers improved performance over existing approaches on both the public benchmark and the in-house dataset. The implementation code has been made available. https://github.com/21wang12/SaTaNet.