Modified anisotropic diffusion and level-set segmentation for breast cancer

Breast cancer is frequent among women and its early diagnosis using thermography is not been widely practiced in medical facilities due to its limitation in classification accuracy, sensitivity, and specificity. This research aims to improve the accuracy, sensitivity, and specificity of breast cancer classification in thermal images. The proposed system is composed of the Least Square Support Vector Machine (LSSVM) to improve the classification and prediction accuracy of breast thermography images using optimized hyperparameters. Multi-view breast thermal images are pre-processed using Gaussian Filtering (GF) with a standard deviation value of 1.4 which is followed by anisotropic diffusion while trying to enhance the image by removing noise. Interested regions are segmented by the level-set segmentation technique, and canny edge detection is applied to the segmented output to limit the amount of data and filter useless information. Texture features are extracted from 1370 healthy and 645 sick subjects fetched from Database for Mastology Research (DBR) which is an online free thermogram database. The features from different views of thermograms are later reduced with a t-test. Significant features are added together to obtain feature vector which produces vectors that are further supplied to the Vector Support Machine that utilizes optimized hyper-parameters for the breast thermogram classification. Compared to the state of art solution, the proposed system increased the accuracy by 9% while sensitivity and specificity get increased by 5.75% and 7.25% respectively. The proposed method focuses on modifying the anisotropic diffusion function and enhancing the segmentation of breast thermograms for classification analysis.