Dual Thresholding based Breast cancer detection in Mammograms

Early stage detection of Breast cancer may reduce the number of cases in which women lost their life. Motivation of proposed work is to automate the process of breast cancer detection which may strengthen the decision of Radiologist. In proposed work, two thresholding techniques are used in parallel to differentiate the pixels of interest from noninterest (less informative like background pixels). One of these techniques is based on Histogram Peak Analysis (HPA), with the objective to choose the peak which maximizes the interclass standard deviation. Second technique is extended version of famous Otsu's thresholding. Morphological operations along with image reconstruction are used for artifact removal and image enhancement. Feature extraction and classification of images in to malignant (cancerous) or non-cancerous is obtained using a convolutional neural network. Mammographic Image Analysis Society (MIAS) database is used for performing experimental results.