Cephalometric measurement plays an essential role in analysis of orthodontic mechanisms and orthodontic treatment design. Landmark detection is the first and most important step of cephalometric measurement. Traditional pure film hand drawing and computer software-aided hand drawing methods are time-consuming and involve considerable subjectivity. Current convolutional neural network-based automatic cephalometric measurements methods only use the positional information of the landmarks; the relative spatial information among the landmarks and the angles formed by baselines are not considered and the priorities of key landmarks are ignored, despite their importance for cephalometric measurement. In this paper, we develop an end-to-end framework, consisting of an encoder-decoder module based on a fully convolutional network and a new module based on relational reasoning. The relative distances among landmarks and the proportions and angles formed by baselines are used to build a new loss function. All data used in this manuscript were collected from the West China Hospital of Stomatology and the data set included 1,005 cephalometric X-ray images. Experimental results show that the proposed model improves key landmark prediction accuracy while maintaining the precision of existing prediction results. The results also show that the relational reasoning network can capture the potential relations of landmarks and further improve the prediction accuracy.

Cephalometric measurement, Deep learning, Relational reasoning, Landmark detection