Early anomaly prediction in breast thermogram by hybrid model consisting of superpixel segmentation, sparse feature descriptors and extreme learning machine classifier

Abstract

The breast thermograms can be used to detect location, physiological condition and vascular state of anomalous breast tissues. Most of the schemes reported in literature use breast tissues as region of interest (ROI) for feature extraction and breast anomaly detection. This paper presents a two-level hybrid method for breast thermogram anomaly detection. In the first stage, suspected region-based ROI segmentation model is developed. For this, thermally adaptive superpixels with spatial and temperature coherency are generated by applying linear iterative clustering on pre-processed breast thermograms. Different temperature regions are integrated by clustering superpixels. In the proposed method first and second highest temperature regions are considered as ROI to cover maximum anomalous regions which also make it robust against pseudo colouring. In second stage, shearlet transform is employed on the segmented ROI to obtain co-occurrence matrix-based feature descriptors. The problem of large coefficients in shearlet decomposition is overcome by selecting effective features using kernel-principal component analysis technique. Extreme Learning Machine classifier is employed on a dataset of thermograms to classify the normal and anomalous thermogram. The obtained performance parameters demonstrate the classification accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F-1 score and area under curve of 95%, 93.33%, 96.66%, 96.55%, 93.54%, 94.91% and 95.11%, respectively. The efficacy of the proposed method is also verified by comparing the results; hence, it can be used for early anomaly detection.