Abstract
Fully automated techniques using convolutional neural networks for cephalometric landmark detection have recently advanced. However, all existing studies have adopted X-rays. The problem of direct exposure of patients to X-ray radiation remains unsolved. We propose a model for detecting cephalometric landmarks using only facial profile images without X-rays. First, the model estimates the landmark coordinates using the features of facial profile images through high-resolution representation learning. Second, considering the spatial relationship of the landmarks, the model refines the estimated coordinates. The estimated coordinates are input into fully connected networks to improve the accuracy. During the experiment, a total of 2000 facial profile images collected from 2000 female patients were used. Experiments results suggested that the proposed method may perform at a level equal to or potentially better than existing methods using cephalograms. We obtained an MRE of 0.61 mm for the test data and a mean detection rate of 98.20% within 2 mm. Our proposed two-stage learning method enables a highly accurate estimation of the landmark positions using only facial profile images. The results indicate that X-rays may not be required when detecting cephalometric landmarks.