Classification of Ultrasound Breast Images Using Fused Ensemble of Deep Learning Classifiers

Abstract

Ultrasound (US) imaging is an affordable, radiation-free screening that has been successfully used for early stage breast cancer screening. Deep learning-based classifiers are currently being used to classify breast cancer. Deep learning requires large amount of dataset for training. However, currently available databases of breast cancer US images are small and the images have tumors of different sizes. Therefore, the deep learning-based classifiers are unable to provide good generalization. To address these challenges, we propose a fusion of three models namely transfer learning, multi-scale and autoencoder. Transfer learning model is based on VGG16 and is used to overcome the issue of limited data. Convolutional autoencoders extract features that can represent even noisy images. We propose a novel multi-scale deep learning model to address learning of US images with tumors of various sizes and shapes. These three models are trained independently and then their classification outputs are fused using differential evolution (DE) algorithm to get the final classification results. The proposed novel fused ensemble of deep learning-based classifiers is evaluated using two publicly available US datasets. Transfer learning, autoencoder, and multi-scale models individually achieve an accuracy of 88%, 85%, and 89% respectively. The fusion of the outputs of the three models using DE algorithm provides a classification accuracy with an accuracy of 93%. The source code available at https://github.com/EbrahimAli1989/Breast-Cancer-classification-.git.