Advancements in Research on the Classification of Benign and Malignant Breast Tumors Utilizing Ultrasound Radiomics and Deep Learning

Abstract

This paper examines the applications of traditional machine learning and deep learning in the analysis of breast ultrasound images for tumor diagnosis and explores recent developments in multimodal imaging. Traditional machine learning efficiently classifies breast ultrasound images through preprocessing, feature extraction, and selection, utilizing classifiers such as support vector machines. However, the design of features is highly dependent and its application scope is limited. Deep learning methods, particularly convolutional neural networks, autonomously extract sophisticated features, demonstrating enhanced classification performance and generalization capabilities. For instance, they achieve diagnostic accuracies exceeding 90% in large-scale datasets, with some studies outperforming clinicians. Moreover, this study highlights that the multimodal analysis strategy, integrating breast ultrasound with shear wave elastography, compensates for the limitations of unimodal images and enhances diagnostic accuracy and reliability, signifying a significant advancement in the technology for early breast cancer diagnosis.