International Journal of Computer Applications (0975 – 8887) Volume 47– No., June 2012 33 Textural Analysis of Spinous Layer for Grading Oral Submucous Fibrosis Rusha Patra Department of Electrical Engineering Indian Institute of Technology Kharagpur Chandan Chakraborty School of Medical Science and Technology Indian Institute of Technology Kharagpur Jyotirmoy Chatterjee School of Medical Science and Technology Indian Institute of Technology Kharagpur ABSTRACT Spinous layer being a sub compartment of surface epithelium of oral mucosa, plays also major role in investigating oral submucous fibrosis (OSF) at its early stage in addition to basal layer. This paper aims to provide a technique that can be used to assist the oral pathologists in grading OSF based on textural information of the spinous layer. The proposed scheme intends to evaluate the textural changes from normal to various grades of OSF. In practice, it comprises the following modules – (a) surface epithelium segmentation, (b) selection of windows on the spinous layer in reference to the basal layer, (c) textural feature extraction and analysis and finally (d) grading. Here the epithelium is segmented using anisotropic diffusion and Otsu’s thresholding. Wavelet based multi-resolution technique is applied to extract 12 textural features from spinous layer. From the statistical analysis, it is observed that 6 features are significant in discriminating normal, OSF with and without dysplasia. Finally, support vector machine (SVM) and Bayesian classifiers are trained with 46 normal, 24 OSF without dysplasia and 20 OSF with dysplasia samples for OSF grading. The result shows that the classification accuracies for both the classifiers (Bayesian = 93.3%, SVM = 96.6%) are comparable, there by emphasizing the significance of texture in oral cancer diagnostics. General Terms Image processing, Computer Aided Diagnosis system. Keywords Oral submucous fibrosis (OSF), spinous layer, anisotropic diffusion, wavelet, statistical test, Bayesian classifier, support vector machine. 1. INTRODUCTION Medical history has been witnessed for increasing oral cancer over recent years and more than 0.3 million new cases of oral cancer are reported each year [1]. A high incidence of oral cancer is mainly due to the late diagnosis of potential precancerous lesions and conditions [2]. Oral submucous fibrosis (OSF) is an insidious chronic progressive precancerous condition of the oral cavity [3] and a large proportion of this precancerous lesion converts to squamous cell carcinoma. Histopathologically OSF is characterized by the concomitant presence of less vascularized collagenous connective tissue with overlying atrophic epithelium in oral mucosa [4]. Oral epithelial dysplasia is the diagnostic term used to describe the histopathologic changes seen in a chronic, progressive and premalignant disorder of oral mucosa. In these cases, the epithelial cells show an altered maturation pattern and cytological changes such as cellular and nuclear pleomorphism, nuclear hyperchromatism, increased nuclear/cytoplasmic ratio, prominent nucleoli, increased mitotic activity, increased intercellular space, loss of epithelium cell cohesion and so on [5]. Due to these changes, the overall architecture of the epithelium has changed. Traditionally, the pathologists use histopathological images of biopsy tissue and examined them under light microscope to detect OSF which is a highly qualitative process. Landini and Rippin were the first who tried to overcome the subjectivity in diagnosing the oral malignancy in 1996. They have proposed histopathological diagnosis of oral premalignancy and malignancy based on the morphological characteristics of the cells and tissues in 2-D section of histopathological images [6]. In 2004, Landini and Othman have proposed an automatic method based on morphological reconstruction to describe the hierarchical architectural characteristics of surface epithelium for oral cancer, dysplastic and healthy conditions [7]. The severity of the atypical changes (cellular, nuclear pleomorphism, tear-drop shaped rete-ridges) and the height in the epithelium to which these changes extend have been studied for the grading of dysplasia [8]. Beyond tissue component features (classically the cell and nuclear morphology), the structural characteristics of tissue (e.g. complexity of tissue profile) has been described, but limited work has been done on statistical description of tissue spatial architecture [9-10]. The textural analysis has been mostly done on sub-epithelial connective tissue (SECT). First attempt has been made to analyze the transmission electron microscopic (TEM) images of sub-epithelial collagen fibers in early and advance stages of OSF using wavelet and neural network based technique [11]. Different wavelet features have been studied to classify progressive stages of OSF [12]. Cells in SECT region are also studied and automated classification of SECT cells using SVM based approach have been performed [13]. In the literature, it has been observed that the classifier based on textural analysis of epithelial tissue for the detection of OSF has been attempted less [14]. It can be noted that the oral cancer initiates from the basal layer, so it is very much important to explore the basal layer as well as the other sub-layers of oral mucosa for studying related morphological and textural features. In view of this, the present paper focuses on the quantification of the change of textural aspect of epithelium for the characterization of normal and different grades of OSF. To quantify the overall texture of the spinous or prickle layer cell, the wavelet and subband decomposition of the selected ROI (region of interest) has been performed followed by the classification of the normal and different grades of OSF. To improve the classification accuracy, the statistical analysis of