An Efficient Deep Learning Based Coarse-to-Fine Cephalometric Landmark Detection Method

Accurate and automatic quantitative cephalometry analysis is of great importance in orthodontics. The fundamental step for to annotate anatomic-interested landmarks on X-ray images. Computer-aided method remains be an open topic nowadays. In this paper, we propose efficient deep learning-based coarse-to-fine approach realize accurate landmark detection. the coarse detection step, train a deformable transformation model by using training samples. We register test images reference image (one image) trained predict landmarks' locations Thus, regions interest (ROIs) which include can located. fine utilize convolutional neural networks (CNNs), detect ROI patches. For each landmark, there one corresponding network, directly does regression landmark's coordinates. considered as refinement or fine-tuning based step. validated proposed public dataset from 2015 International Symposium Biomedical Imaging (ISBI) grand challenge. Compared with state-of-the-art method, not only achieved comparable accuracy (the mean radial error about 1.0-1.6mm), but also largely shortened computation time (4 seconds per image).