Segmentation of Breast Cancer Masses in Mammography Images Using Deep Convolutional Neural Network (DCNN)

Mammography is one of the most important and effective ways to detect breast cancer, especially in the early stages of the disease. However, due to the complexity of breast tissue, the similarity between cancerous masses and natural tissues, the different sizes and shapes of masses, and the use of low-density X-ray radiation, mammogram images often have poor quality. Therefore, detecting lesions, especially in the early stages, is a challenging task. In this study, we address the improvement of breast cancer mass segmentation in mammography images. Accurate mass segmentation on mammograms is an important step in computer-aided diagnosis systems, which is also a challenging task because some mass lesions are embedded in natural tissues and have weak or ambiguous margins. The proposed method in this study presents an improved algorithm for segmenting cancerous masses in mammography images using a Deep Convolutional Neural Network (DCNN), which ultimately leads to mass classification into benign and malignant classes. Deep convolutional neural networks extract high-level concepts from low-level features, and are appropriate for handling large volumes of data. In fact, in deep learning, high-level concepts are defined by low-level features. Segmentation based on the proposed method with preprocessed images achieves more accurate delineation in high-resolution images, and ultimately improves the accuracy and sensitivity of mass tissue separation in the breast. In this study, we used three different architectures for deep convolutional neural networks. The proposed DCNNs were validated on mammography images of INbreast dataset. The reliability of the system's performance is ensured by applying 5-fold cross-validation. The proposed method has been evaluated based on accuracy, precision, sensitivity, and ROC criteria. The results obtained with an accuracy of 97.76% for the third proposed deep model indicate an improvement in the accuracy of the diagnosis as well as a reduction in the cost of the diagnostic process. Results showed that our proposed algorithm is significantly more accurate than other methods due to its deep and hierarchical extraction.