Locating Cephalometric Landmarks with Multi-Phase Deep Learning

Abstract

Cephalometric analysis has long been one of the most helpful approaches in evaluating cranio-maxillo-facial skeletal profile. To perform this, locating anatomical landmarks on an X-ray image is a crucial step, demanding time and expertise. An automated cephalogram analyzer, if developed, will be a great help for practitioners. Artificial intelligence, including machine learning is emerging these days. Deep learning is one of the most developing techniques in data science field. The authors attempted to enhance the accuracy of an automated landmark predicting system utilizing multi-phase deep learning and voting. To guarantee objectivity, an open-to-the-public dataset, cephalometric images accompanied with coordinate values of 19 landmarks, were used. A regressional system was developed, consisted with convolutional neural networks of three phases. First phase network was to determine approximate position of each landmark, inputting whole area of compressed original images. Five secondary networks were to narrow down the area, based on the first phase prediction. Third phase networks were trained by small areas around respective landmarks, with original resolution. Third phase prediction with voting was done inputting 81 shifted areas. Successful detection rates improved as the phase advances. Voting in third phase improved successful detection rate. In comparison with previously reported benchmarks, using the same dataset, proposed system marked better results. Within the physical limitation of memory and computation, multi-phase deep learning may be a solution to deal with large images.