Applications of a novel deep neural network to the classification of liver steatosis and breast lesions in ultrasound images

Abstract

Ultrasound is one of the most commonly used imaging modality in clinical practice. A computer-based approach is required for the non-invasive detection of chronic liver diseases or breast cancers. In particular, breast cancer is a substantial public health concern, and prompt detection and classification are crucial for the effectiveness of treatment. In this study, a divergence-based feature extractor (DivFE), a new deep learning model focusing on the identification of features, is used. Image features are extracted with convolutional neural networks (CNN) trained with Walsh vectors and the classification process is carried out with minimum distance network (MDN). In the literature, it is seen that ultrasound images of breast and liver diseases are successfully classified using deep neural networks (DNN). In this study, the same images were classified with high classification accuracy using a smaller number of nodes compared to the DNNs in the literature. To demonstrate the advantages of the DivFE, four widely used datasets were employed: three breast cancer datasets (Datasets I, II, and III) and one liver steatosis dataset (Dataset IV). Dataset I, Extended Dataset (I+II), Dataset III and Dataset IV, classification success rates of 100%, 97%, and 91%, and 100% were respectively achieved by using the DivFE with a small number of nodes. It is seen that classification accuracies obtained in the literature were achieved by using a new, small-sized DNN.