Ensemble deep learning system for early breast cancer detection

This paper proposes an ensemble deep learning system for the early detection of breast cancer. Unlike traditional ensemble learning that processes the whole image, the proposed system processes only the Suspected Nodule Regions (SNRs), extracted using an optimal dynamic thresholding method, where the threshold varies corresponding to the details of each input image. The SNRs affords better performance and offer the ability to detect small size nodules. The ensemble is composed of four transfer learning Convolutional Neural Networks (CNNs) (i.e., AlexNet, ResNet-50, ResNet-101, and DenseNet-201). A binary Support Vector Machine (SVM) follows each CNN model to provide either a malignant or a benign score. To obtain the final system decision, the first-order momentum is derived over the four binary outputs of the ensemble CNNs and the final decision is guided by the four CNNs’ training accuracies. The proposed ensemble fusion scheme, tested on Region of Interest (ROI) images, using the public digital medical images Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM), achieves an accuracy of 94% to distinguish between Malignant (M) and Benign (B) classes and of 95% to distinguish between Malignant Mass (MM) and Benign Mass (BM) nodules. Comparison with related methods on the same data confirms the accuracy and advantages of the proposed ensemble system.