Document Type : Full Research Paper

Authors

• Raheleh Kafieh ¹
• Alireza Mehri Dehnavi ²
• Saeed Sadri ³
• Seyed Hamid Raji ⁴

¹ MSc Graduated, Department of Biomedical Engineering, Medical School, Isfahan University of Medical Sciences

² Associate Professor, Department of Biomedical Engineering, Medical School, Isfahan University of Medical Sciences

³ Associate Professor, Electrical & Computer Engineering School, Isfahan University of Technology

⁴ Assistant Professor, Department of Orthodontics, Dentistry School, Isfahan University of Medical Sciences

Abstract

Cephalometry is the scientific measurement of head dimensions to predict craniofacial growth, plan treatment and compare different cases. There have been many attempts to automate cephalometric analysis with the aim of reducing the time required to obtain an analysis, improve the accuracy of landmark identification and reduce the errors due to clinician subjectivity. This paper introduces a method for automatic landmark detection on cephalograms. We introduced a combination of model­-based methods and neural networks on cephalograms. For this purpose, first some feature points were extracted using a nonlinear diffusion filter and Susan Edge Detector to model the size, rotation, and translation of skull. A neural network was used to classify the images according to their geometrical specifications. Using learning vector quantization (L VQ) for every new image, the possible coordinates of landmarks were estimated. Then a modified active shape model (ASM) was applied and a local search to find the best match to the intensity profile was used and every point was moved to get the best location. Finally, a sub-image matching procedure was applied to pinpoint the exact location of each landmark. In order to evaluate the results of this method, 20 randomly selected images were used with a drop-one-out method. Each image had a dimension of about 170x200 mm, digitized in 100 dpi (4 pixel == 1mm). On average, 24% of the 16% landmarks were within 1mm of correct coordinates, 61 percent within 2 mm, and 93 percent within 5 mm. the proposed method in this study has had a distinct improvement over the other proposed methods of automatic landmark detection. 

Keywords

• Cephalometry
• Active shape model (ASM)
• Non-linear diffusion
• Neural Networks
• Sub-image matching

Main Subjects

• Biomedical Image Processing / Medical Image Processing