Five Classification of Mammography Images Based on Deep Cooperation Convolutional Neural Network

Abstract

Mammography is currently the preferred imaging method for breast cancer screening. Masses and calcification are the main positive signs of mammography. Due to the variable appearance of masses and calcification, a significant number of breast cancer cases are missed or misdiagnosed if it is only depended on the radiologists’ subjective judgement. At present, most of the studies are based on the classical Convolutional Neural Networks (CNN), which uses the transfer learning to classify the benign and malignant masses in the mammography images. However, the CNN is designed for natural images which are substantially different from medical images. Therefore, we propose a Deep Cooperation CNN (DCCNN) to classify mammography images of a data set into five categories including benign calcification, benign mass, malignant calcification, malignant mass and normal breast. The data set consists of 695 normal cases from DDSM, 753 calcification cases and 891 mass cases from CBIS-DDSM. Finally, DCCNN achieves 91% accuracy and 0.98 AUC on the test set, whose performance is superior to VGG16, GoogLeNet and InceptionV3 models. Therefore, DCCNN can aid radiologists to make more accurate judgments, greatly reducing the rate of missed and misdiagnosis.