CLASSIFICATION OF ORAL LESIONS USING COLOR IMAGES

Abstract

The work revolves around the classification of oral lesions using color images. Oral cancer can be treated successfully if diagnosed at an initial stage. Present work has focussed on development of Clinical Decision Support (CDS) system for qualitative and quantitative analysis of color images of oral cavity. The first part of analysis has focussed on interpolation of color images to an extent where high resolution image assists a medical expert to precisely observe the texture of malignant or benign area more lucidly. Four different linear and non-linear interpolation algorithms have been implemented on color images of oral cavity and their results have been compared on the basis of visual clarity and time consumed. All the algorithms have been implemented in a CDS system. The second part of analysis has proposed a technique which uses color images of oral cavity as an input data to predict malignancy at an early stage of disease. Image segmentation of a manually selected region has been performed by using a semi-automatic technique i.e. Active Contour without Edges and patches have been selected from segmented region. Based on intensity the first order features have been evaluated directly for each patch. GLCM has been a conventional method of evaluating second order statistical texture features. In present work RABGLDCM has been used which meliorated the texture information provided by GLCM. - Feature selection is a significant part of any classification problem and algorithms of feature selection galore in literature. Due diligence has been paid in performing this crucial task and • two separate approaches of feature selection i.e. filter approach and sequential feature selection (SFS) have been followed. Five different sets of features were prepared depending upon their discriminative ability and inter-feature dependency. In the end binary classification has been performed using Support Vector Machines (SVM) using five different sets of features which were obtained after feature selection process. The performance of the classifiers has been perspicuously represented in their sensitivity, specificity and accuracy. RABGLDCM combined with SVM and SFS yielded best results. Database comprising of 50 benign and 20 malignant case color images of oral cavity, used in the present work has been collected from Cancer Research Institute under Himalayan Institute of Medical Sciences, Dehradun.