Unsupervised Learning for Breast Abnormality Detection using Thermograms

Breast cancer is the most diagnosed cancer in women globally. Thermography, an adjunctive tool to mammography, may be used for early detection of breast abnormality. The existing supervised learning techniques in the literature for abnormality detection using breast thermogram images require a big training set of labeled samples for adequate learning. However, labeling a large dataset is a resource-intensive and time-consuming process. Therefore, this work proposes a novel unsupervised learning methodology that employs a gradient-based local feature descriptor for abnormality detection using breast thermograms. Different measures of similarity (or distance) between the histogram of oriented gradients features (HoG, a gradient-based local feature descriptor) of the left and the right breast of an individual are used to identify the asymmetry between both breasts. The distances between the HoG features of both breasts are clustered using an unsupervised clustering algorithm called the
$K$
-means algorithm. The cluster with a greater centroid value (higher mean of distance values) indicates asymmetry and is labeled as abnormal. The red-plane of a thermogram containing the regions of higher temperature is extracted for enhanced gradient information and thereby, for enhanced detection of breast abnormality. This work pioneers the application of asymmetry-based unsupervised learning for the detection of breast abnormality in thermograms and achieves an accuracy of 86.52%.