Accurate Hotspot Segmentation in Thermal Breast Images with Gaussian Mixture Model Superpixels

Thermography is a non-invasive and promising clinical method widely being used in screening of breast cancer. Breast thermal images show the heat distribution, depicting the manifestation of cancerous lesions with hot regions. Accurate segmentation of these hotspots is very vital in determining the severity of the lesions, localization and treatment regimen. In this work, color segmentation based on superpixel scheme is proposed to extract the hotspots from breast thermal images. Here, we employ the most recently proposed Gaussian mixture model superpixels, featuring semantically defined boundaries to segment the hotspots with ease. The proposed system tested with a standard dataset exhibits best performance, evaluated with various statistical metrics. It demonstrates very high segmentation accuracy around 99.84%. Visual analyses and statistical comparisons with ground truth images establish the superiority of proposed system over existing methods. Proposed method could increase diagnostic accuracies of breast cancer and increase the abnormality detection in early stage.