A Deep System for Breast Tumor Classification from Mammograms

Abstract

Medical imaging has continued to evolve and uses increasingly precise and efficient technologies for processing medical images and extracting relevant information. However, in contrast to this development, the analysis of breast cancer remains topical research that is very delicate to address despite the various screening techniques available to predict this disease. In this paper, a novel system based on deep learning is proposed to detect and classify breast tumors in mammograms into benign and malignant. Two approaches have been adopted in this study according to two publicly available mammographic datasets for research of different sizes. In the first approach, we have built a model on the Mammographic Image Analysis Society (MIAS) dataset which is a small set. To make this dataset sufficient to build a predictive model over, the first approach is passed through three steps, which are as follows: First, the Regions Of Interest (ROI) are extracted from each abnormal region of mammographic images of MIAS set, then, Roi-patches are generated from each ROI obtained in the first step, finally, ROI-patches are classified as benign or malignant using a pre-trained Convolutional Neural Network (CNN) architectures. While the second approach uses the cropped images extracted from the Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM) which are considered sufficient to build a model over it. Data preprocessing as well as Data augmentation has been applied on both MIAS and CBIS-DDSM datasets. Experiments are evaluated through transfer learning techniques based on Convolutional Neural Networks (CNNs) architectures: Dense-Net121, ResNet50, VGG16, and mobileNetV2 to check their accuracy scores. Experimental results are satisfactory showing the proposed approaches’ ability to distinguish between breast tumor benign and malignant from Mammograms with an overall sensitivity of 91% and accuracy of 90%.