CephaNet: An Improved Faster R-CNN for Cephalometric Landmark Detection

Abstract

In orthodontic treatment, automatic cephalometric landmark detection is essential for clinical diagnosis, planning, and research. In accordance with the features of the cephalograms and the distribution of cephalometric landmarks, this paper proposes the first Faster R-CNN based method, CephaNet, for cephalometric landmark detection. In CephaNet, we design a multi-task loss for reducing intra-class variations and adopt the multi-scale training strategy to improve detection accuracy of small landmarks. For removing abnormal detected landmarks (superfluous or undetected landmarks), we present a two-stage repair strategy. Firstly, we construct a 2D undirected graph according to the distribution of landmarks in training data, and then we adopt ’max-confidence’ and Laplacian transformation to remove abnormal landmark-s. CephaNet obtains the state-of-the-art performance in the public dataset, and its detection accuracy is about 6% higher than other methods in the clinically accepted 2-mm range. The results demonstrate CephaNet is effective in cephalometric landmark detection.