Automatic cephalometric landmarks detection

Abstract

Cephalometric landmarks are required for Cephalometric analysis that is used in orthodontic analysis and human facial structure diagnosis. Thus, in order to reduce time and needs of expert orthodontists, many attempts have been proposed to use computers instead of manual procedure to acquire those landmarks. Our proposed method is to create two non-deformable general models by Procrustes analysis, one for the upper part and the other, the lower, of a human facial structure. By placing and aligning the general models on a test image, considering some anatomical conditions, regions of interest (ROI’s) of the target landmarks are automatically identified. Our algorithm is evaluated by leave-one-out method, it is found that the average success rates of finding the landmarks within the +/- 2 SD ROI’s are 85.84% and 83.33% for the upper and the lower models respectively. For error calculation, an expert orthodontist has been asked to locate landmark twice for some images in order to evaluate the expert accuracy. We found that our methodology gave the accuracy of some landmarks as similar as the expert. Moreover, some detection gave better results than an expert. From 13 landmarks, 8 landmarks were detected within 3 mm. error, while only 2 landmarks were detected with more than 5 mm. error.