Abstract
Cephalometric analysis is very essential for the patient having dentofacial and craniofacial deformities. The manual localization of the cephalometric landmarks is also important and critical for the observer that is required to be performed by the orthodontics only. The manual localization is the time consuming and the tedious task for the observer. Therefore, we proposes a method to automatically detect cephalometric landmarks on lateral cephalometric x-ray image. The proposed method is a deep learning approach where centroid based registration was performed on the same size images then ResNet50 was applied on the different patches which were made based on the geometrical position of the landmarks. Total ten patches were made for the 19 landmarks. The average landmark detection rate during the testing was achieved as 90.39% and 92.37% under 2-mm and 3-mm error respectively for testset1 database. The average landmark detection rate during the testing was achieved as 82.66% and 84.53% under 2-mm and 3-mm error respectively for testset2 database. The average mean error and standard deviation on testset1 was found as 1.23 mm and 0.73 respectively and average mean error and standard deviation on testset2 was found as 1.37 mm, and 0.88 respectively. The proposed method was compared with the state-of-the-art methods and found the improved results in terms of successful landmark detection rate under 2-mm. The results were found very promising and the proposed method may be helpful to use in clinics further.