Improvement of breast density classifier based on CNN features extraction and SVM in mammogram images

Screening mammography has been clinically practiced as a common method for monitoring any potential breast diseases, especially in denser breasts among women. The advancement in medical imaging technology proved that the integration of artificial intelligence had given a significant impact on the breast screening process. However, due to various demographic patient backgrounds in clinical profile and non-standardized configurations of the developed intelligence models, such application is incapable of being applied by a health practitioner. Commonly, the deep learning method trained on a single network classifier has resulted in lower performance accuracy. With the motivation to improve the performance of classifying the dense breast, this paper proposes an improved classification model using deep learning approach of Convolutional Neural Network (CNN) and Support Vector Machine (SVM) employed on different sets of publicly established mammogram images of craniocaudal (CC) and medio-lateral (MLO) views. In this study, pre-trained CNN models, namely GoogleNet, ResNet50, ResNet101 and AlexNet, are used with SVM as a classifier. The proposed method for the classification of breast density region is superior to the existing methods as indicated from model performance quantitative results of accuracy, precision and area under the curve (AUC) of its receiver operating characteristic (ROC) curves. Significant improvement in model performance has been obtained using ResNet50 and GoogleNet with SVM classifier with > 94% accuracy and AUC > 0.95. Furthermore, the model is proved to have good feature extraction capabilities to work for various breast density images that can be further explored into detecting malignancy from screening mammogram images.