Developing a toolbox for clinical preliminary breast cancer detection in different views of thermogram images using a set of optimal supervised classifiers

Abstract

A full automatic technique and a user-friendly toolbox are developed to assist physicians in early clinical detection of breast cancer. The database contains different degrees of thermal images obtained from normal or cancerous mammary tissues of patients with mean age of 42.3 years (SD:±10.50), whose sympathetic nervous systems were activated with a cold stimulus on hands. First, ROI was determined using full automatic operation and the quality of image was improved. Then, some features, including statistical, morphological, frequency-domain, histogram, and GLCM, were extracted from segmented right and left breasts. Subsequently, to achieve the best feature space for decreasing complexity and increasing accuracy, feature selectors such as mRMR, SFS, SBS, SFFS, SFBS, and GA were used. Finally, for classification and TH labeling, supervised learning techniques such as AdaBoost, SVM, kNN, NB, and PNN, were applied and compared with each other to find the most suitable one. The experimental results obtained on native database showed the mean accuracy of 88.03% for 0-degree images using combination of mRMR and AdaBoost and for combined 3 degrees using combination of GA and AdaBoost. The maximum accuracy obtained from all degrees and their combinations before and after ice test was nearly 100%.