Suspicious Region Segmentation Using Deep Features in Breast Cancer Mammogram Images

One of the most challenging processes in the image processing field is segmentation, especially region of the breast and Pectoral Muscle segmentation in mammogram images due to the presence of artifacts, homogeneity between the Region of Interest and pectoral muscle, also because of low contrast along the boundary of breast region. Moreover, similarities in texture between the texture of Region of Interest and pectoral muscle and irregularities in Region of Interest. The main objective of this work is to propose a model-based deep convolutional neural network feature to segment the Region of Interest from the pectoral muscle. The pre-processing stage has been employed in the first step. Morphological operations with Otsu's thresholding are used to remove artifacts and labels in images. After that, Wavelet Transform has been exploited to remove noise, and histogram equalization is employed to ameliorate image contrast as a last step of pre-processing. Then, Artificial Neural Network and Support Vector Machine classifiers are trained based on deep convolutional neural network features to estimate the Region of Interest from the pectoral muscle. This work was evaluated by employing 322 mammogram images from the mammographic image analysis society (mini-MIAS). Accordingly, the Region of Interest segmentation method achieved an accuracy of 95.34%.