Cephalometric Landmarks Detection using Fully Convolutional Networks

Abstract

In dentistry, quantitative cephalometry plays an essential role in the practice of medical care for patients. In this thesis, an automated landmark detection model is proposed using FCN (fully convolutional networks) with internally residual connections. The FCN model was trained to output an archery target shape heatmap when an image patch near the landmark was input. The image patches used for training were positioned and sized based on training data, and augmentation was performed. The cephalogram were used for training and testing used a publicly available datasets. SDR(Success detection rate) was used to evaluate the results. The trained models were evaluated using a test set and compared with previous studies. As a result, landmarks were detected with better accuracy than previous studies. The FCN model showed the potential for accurate landmark detection.