Deep Representation for the Classification of Ultrasound Breast Tumors

Abstract

An automated classification of ultrasound breast tumor is a vital step for the early prevention of abnormal breast cells. In general, radiologists manually handle this procedure, but manual analysis performed by individual poses a problem of consistency depending on the experts. One of the standardized alternatives was to apply automated deep learning method in this field. In fact, majority ideas in literature are dominantly based on the supervised learning framework, but even such methods have still failed to present promising discrimination performance. In this work, we assume that unsupervised learning still can be a potential option and beneficial attribute that enables to accelerate discrimination is inherent in it. Hence, we present a deep representation for the ultrasound breast data utilizing two types of independent supervised and unsupervised network to reconstruct the principal features, while the volume of supervised features is set to be minimum and the volume of unsupervised is the maximum. Specifically, we adopted pretrained Resnet34 as a supervised network, and a convolutional autoencoder (CAE) was designed for the unsupervised network. Each representation vector is combined into a single vector, and the generated vector is given to the support vector machine as an input for the final discrimination. The results are verified that the proposed method shows far better performance compared to several conventional deep learning methods and the single use of each method. The value of accuracy, sensitivity and specificity are obtained by 88.18%, 85.25% and 100.00% respectively.