DIGITAL INFRARED IMAGING FOR BREAST CANCER DETECTION USING SEQUENTIAL MINIMAL OPTIMIZATION, KERNEL LOGISTIC REGRESSION AND MULTILAYER PERCEPTRON

Abstract

Breast cancer is the main reason for women's death. In Egypt, breast cancer is considered to be the most common cancer among females. Death percentage by breast cancer is up to 37.7% out of 12.621 new cases in 2008. Breast cancer starts when cells in the breast begin to grow out of control. However, there are many techniques help to discover the cancer in a more safe way. One of them is Digital Infrared Imaging (thermography); it is based on the metabolic activation and vascular circulation in both pre-cancerous tissue and the cancerous one. Radiographic images obtained from thermography equipment are one of the most f techniques that used for helping in early detection of breast cancer. This thesis proposes a method for breast cancer detection that uses image processing techniques. These techniques are applied to 142 breast digital thermal images; 77 of them are normal images and 65 are abnormal ones. Matlab, is used for detecting region of interest (ROI) and feature extraction. In addition, Weka is used for classification and feature selection. In the (ROI) extraction phase, active contour techniques are used, and then 72 statistical and textural features are extracted and used to feed classifiers. It is worth mentioning, there are three different classifiers. These classifiers are Sequential minimal optimization (SMO), kernel logistic regression, and Multilayer perception and their accuracy had reached to 99.29%, 98.59%, 96.4%. Afterwards it comes to the feature selection phase, at which the best dominant features are selected to minimize the processing time and effort. Accordingly, there are two types of wrapper feature selection methods are called best first and Greedy Step wise. These methods are used to minimize the number of features. Finally, only 6 features were selected, which gave us an accuracy of 99.29% using Sequential Minimal Optimization (SMO), 99.29% using Kernel Logistic Regression, and 97.18 using Multilayer Perceptron.