Ultrasound Elastography Imaging for Breast Lesions Classification Using Deep Transfer Learning
Monday, April 8, 2024
2:48pm – 2:55pm
Location: 412
Authors: Huimei Cheng, Apple Inc. Yi-Hong Chou, Taipei Veterans General Hospital Chui-Mei Tiu, Yeezen General Hospital Hsiao-Chuan Liu, University of Southern California
Ultrasound elastography imaging is a powerful tool for providing substantial mechanical information about tumor stiffness to radiologists during the diagnosis process. Typically, it uses intensity information in imaging to represent the stiffness of biological tissues in commercial ultrasound instruments. However, elastography images may sometimes fail to accurately reflect the mechanical properties of biological tissues because their stiffness may not truly correspond with their intensity representation. Although deep learning shows promise in potentially improving accuracy, ultrasound elastography images often suffer from a lack of sufficient data for training models. In this study, we propose the use of deep transfer learning to accurately differentiate between malignant breast elastography images and benign lesions, and we compare the results with a machine learning method. We report that the information contained within breast elastography images, along with their morphologies, can be valuable features for enhancing accuracy, compared to using intensity information alone, and deep transfer learning exhibits superior performance in ultrasound elastography classification, particularly when confronted with limited sample sizes.