Efficient and Stable Cephalometric Landmark Localization Using Two-Stage Heatmaps’ Regression

Abstract

Cephalometry is commonly used as an integral component in orthodontic treatment and visual treatment objects( VTO). Accurate and stable tissue landmarks localization is an essential precondition for orthodontic diagnosis and treatment. In practice, expert manual annotation is often time-consuming and subjective. Recent works based on deep learning have two approaches for automatic cephalometric assessment, namely multi-model-based and end-to-end model-based methods, leading to substantial differences in landmarks localization accuracy. The multi-model-based methods predict each landmark with higher localization accuracy, but they are complex and their parameters have to be adjusted iteratively. The end-to-end models are both lightweight and efficient and able to predict all landmarks at once, but their localization accuracy is significantly lower than that of the multi-model-based methods. This paper aims at rethinking the different nature among models and present an effective solution for the low accuracy end-to-end model. On the one hand, we design a two-stage network architecture to improve the model representation ability. We found that the input image with high resolution and multi-stage strategy can effectively improve the localization accuracy of the model. Therefore, the former stage employs an HR-Net to output multiple-scale feature maps in parallel from high-resolution radiography images. The latter stage further refines the landmark location from the generated features. On the other hand, thanks to its spatial representation of landmark location, the network-based heatmap prediction has become a mainstream method for regressing the landmarks coordinates thanks. But none so far has the heatmaps decoding method been studied in cephalometry. Thus, we propose an improved heatmaps decoding method that effectively reduces the landmarks localization error during training. Furthermore, experimental results show that our approach outperforms previous related works on a public dataset and yields state-of-the-art accuracy of cephalometric landmarks detection.