Detection of breast cancer tumours based on feature reduction and classification of thermograms

The patients having malignant breast tumours if detected in early stage have a better chance of survival. It is observed that the analysis of the texture features of the breast thermograms helps in providing the right information for diagnosis to a greater extent. In this study, the breast thermograms of 56 subjects having temperature recordings available at Database Mastology Research (DMR), visual labs are considered. Further, the texture features in the Gray level Run Length Matrix (GLRLM) and Gray level Co-occurrence Matrix (GLCM) are extracted from these images. The correlation of features gives a linear relationship between the variables that help to analyse the quantitative relationship between the variables. The features are selected by using unsupervised feature reduction techniques, i.e. Principal Component Analysis (PCA) and Autoencoder (AE). The features selected are observed to be relevant in detecting the abnormality between healthy and unhealthy breast. Different classifiers viz. support vector machine, decision tree, random forest, K-NN, linear Regression, and fuzzy logic are then applied to the selected features for detecting the presence of malignancy in breast. Among all the classifiers, Random Forest (RF) with PCA has been observed to yield an accuracy of 95.45% in classifying the benign and malignant tumours.

Automated summary

This study focuses on analyzing the texture features of breast thermograms to detect malignant breast tumors early. By using unsupervised feature reduction techniques like PCA and AE, relevant features are selected for classification with different classifiers, with Random Forest (RF) observed as the most accurate classifier.