Automatic localization of craniofacial landmarks using multi-layer perceptron as a function approximator I. El-Feghi * , M.A. Sid-Ahmed, M. Ahmadi Department of Electrical and Computer Engineering, University of Windsor, Windsor, ON, Canada N9B 3P4 Received 18 May 2005; received in revised form 9 September 2005 Available online 2 December 2005 Communicated by Prof. G. Sanniti di Baja Abstract There are 20–30 visible landmarks in the lateral X-ray skull that are used by orthodontists in what is known as cephalometric eval- uation. The evaluation assists in the diagnosis of anomalies and in the monitoring of treatments. Currently, this process is carried-out manually by outlining the soft and bonny tissues of the skull then locating the landmarks on line crossings. This can take an experienced orthodontist up to 20 min. The process is tedious, time consuming and subject to human error. In this paper, we propose a system for automatic localization of cephalometric landmarks using Multi-Layer Perceptron (MLP). Image processing techniques are utilized to extract features representing rotation, scale and distances from the outer edges of the skull image. Features from manually labeled images are used as inputs to train the MLP. After training, the MLP is used to estimate the location of the landmarks on targeted images based on knowledge obtained on the training stage. Results obtained by testing the algorithm on images which are not seen by the MLP during training, show an improvement over previously reported techniques. Ó 2005 Elsevier B.V. All rights reserved. Keywords: Cephalometric; Cephalometric evaluation; X-ray image; MLP; Craniofacial 1. Introduction Cephalometry is defined as the scientific measurement of the head usually on digitized X-ray of a skull. This mea- surement is performed by orthodontists based on the loca- tion of a set of agreed-upon points known as craniofacial landmarks. There are 20–30 landmarks visible on the X- ray of a human skull, which are used by orthodontists in what is known as cephalometric evaluations. Location and definition of the most commonly used landmarks, as defined in (Rakosi, 1982), are shown in Fig. 1. Once the landmarks are located, several linear and angular measure- ments are performed to assess the treatment, evaluate trea- ted cases, plan the treatment or compare the measurements with norms of populations. Currently X-ray landmarking is performed manually by first tracing the X-ray, and secondly, locating landmarks on line intersections or geometrical line shapes. This pro- cess requires about 15–20 min per patient and it is subject to human error. A computerized system that will carry out this tedious task is obviously needed and can be useful not only for cephalometric evaluation but also for keeping and organiz- ing dental records. Automated cephalometry has been subject to research for many years and has been attempted by several indepen- dent researchers with varying degrees of success. Previous work can be divided into two categories; in the first cate- gory, the positions of the landmarks are determined based on a set of predefined geometrical properties of lines, line intersections and exterior boundaries as in (Le ´vy-Mandel 0167-8655/$ - see front matter Ó 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.patrec.2005.09.019 * Corresponding author. E-mail addresses: idrise@uwindsor.ca (I. El-Feghi), ahmed@uwind- sor.ca (M.A. Sid-Ahmed), ahmadi@uwindsor.ca (M. Ahmadi). www.elsevier.com/locate/patrec Pattern Recognition Letters 27 (2006) 544–550