Statistical Analysis Based Feature Selection for Detection of Breast Cancer Using Thermograms

Abstract

Breast Cancer is one of the most prevalent diseases among women. Its early diagnosis helps to increase the survival rate. Among many modalities, thermography is considered to be an early diagnostic procedure, which depicts the temperature values of the hot regions and further provides scopes in locating the tumor. In this work, features from Gray Level Co- occurrence Matrix (GLCM) and Gray Level Run Length Matrix (GLRLM) are extracted using the breast thermograms. Dimensionality reduction technique i.e., Autoencoder is applied to the extracted features. It gives the non-linear pixel intensities of the breast thermograms. Further, the reduced feature set is directed towards the statistical analysis of the features with three different methods viz. Filter, Wrapper and Embedded methods on the breast thermograms for the selection of best features set. Random Forest and Decision tree based classification algorithms are further applied for the features selected using three statistical tests. Among both the classifiers, Random forest with Recursive feature Elimination method gives a better performance in detecting the tumors between healthy and unhealthy breasts, giving an accuracy level of 81.63%.