Article Preview
Top1. Introduction
In the last decades, medical imaging techniques have been applied to scan human body for clinical purposes (diagnoses and surgical planning) and medical science studies like anatomy and physiology. The area incorporates imaging technology based on radiography (projection or fluoroscopy), nuclear medicine, endoscopy, microscopy (for human pathological investigations), tomography, magnetic resonance imaging, among others (Suetens, 2002). From the viewpoint of computer science, each one of these modalities needs specific algorithms and mathematical methods for automating the process of image analysis (Suri, 2005).
The Panoramic X-Ray images are obtained through projection radiography devices and are very popular as a first tool for diagnosis in odontological protocols (Figure 1). The panoramic X-ray is a cheap and very useful tool for dentist diagnosis. It has a low level of radiation, is comfortable and very quickly for the patient to have taken. The panoramic X-ray shows the dentist a patient's nasal area, sinuses, jaw joints, teeth and surrounding bone. It can reveal cysts, tumors, bone irregularities, among other problems. So, automating the process of analysis of such images is an important application. However, it is a difficult task due to the structures overlapping and texture patterns commonly observed in such images.
Figure 1. A tipical panoramic x-ray image
Tooth segmentation from dental X-ray films is an essential step for automating diagnosis as well asforensic procedures like postmortem identification (Said et al., 2006). For these applications, the segmentation step must be followed by features extraction. Approaches in image segmentation can be roughly classified in: (a) Contour Based methods, like snakes and active shape models (Suri, 2005); (b) Region based techniques (Zhou & Abdel-Mottaleb, 2005); (c) Optimization approaches (Fukunaga, 1990) ;(d) Clustering methods, like k-means, Fuzzy C-means, Hierarchical clustering (Jain et al., 1999) and (e) Thresholding methods (Sezgin & Sankur, 2004).
Snake models are 2D deformable models that can integrate image data, an initial estimated, prior shape information (active shape model) in a single extraction process (Black & Yuille, 1993). Region-based methods attempt to group regions following some homogeneity criterion. Optimization approaches rely on an objective function, that combines the input image and constraints, and some optimization process for separating foreground from the background in an image. Clustering methods (in general) apply statistics and machine learning to partition N pixels into K clusters (
). Thresholding techniques compute a global threshold to distinguish objects from their background. They are low computational cost and effective tools to segment objects in a scene. Also, it is simple to implement these techniques in a computer language (Kanungo at al., 1994).
There are few researches dedicated to the specific problem of dental radiograph image segmentation. Projection histograms, thresholding, mathematical morphology, contrast enhancement are common techniques applied in this area (Jain & Chen, 2004; Zhou & Abdel-Mottaleb, 2005; Nomir & Abdel-Mottaleb, 2005). Recently, it was proposed a work that also applies zonal masks in the Fourier domain to address overlapping (Lira et al., 2009), which is a very common problem in panoramic X-ray images, as we can see in Figure 1.