Artificial intelligence-based automated breast ultrasound radiomics for breast tumor diagnosis and treatment: a narrative review

Abstract

Breast cancer (BC) is the most common malignant tumor among women worldwide, posing a substantial threat to their health and overall quality of life. Consequently, for early-stage BC, timely screening, accurate diagnosis, and the development of personalized treatment strategies are crucial for enhancing patient survival rates. Automated Breast Ultrasound (ABUS) addresses the limitations of traditional handheld ultrasound (HHUS), such as operator dependency and inter-observer variability, by providing a more comprehensive and standardized approach to BC detection and diagnosis. Radiomics, an emerging field, focuses on extracting high-dimensional quantitative features from medical imaging data and utilizing them to construct predictive models for disease diagnosis, prognosis, and treatment evaluation. In recent years, the integration of artificial intelligence (AI) with radiomics has significantly enhanced the process of analyzing and extracting meaningful features from large and complex radiomic datasets through the application of machine learning (ML) and deep learning (DL) algorithms. Recently, AI-based ABUS radiomics has demonstrated significant potential in the diagnosis and therapeutic evaluation of BC. However, despite the notable performance and application potential of ML and DL models based on ABUS, the inherent variability in the analyzed data highlights the need for further evaluation of these models to ensure their reliability in clinical applications.

Keywords: artificial intelligence; automatic breast ultrasound; breast; breast tumor; deep learning; machine learning; radiomics.

Figure 1

Flow diagram of study selection.

Figure 2

A set of 3-D ABUS images generated by Invenia ABUS.

Figure 3

The radiomics workflow for ABUS imaging integrates two distinct methodologies. Traditional machine learning relies on manual processes including image acquisition, feature extraction and selection, followed by model construction to complete tasks. Deep learning primarily employs convolutional neural networks (CNN) composed of three core components: convolutional layers for local feature extraction, pooling layers for dimensionality reduction, and fully connected layers that map extracted features to output layers for final task execution. SVM, Support Vector Machine; LASSO, Least Absolute Shrinkage and Selection Operator; RF, Random Forest; ROC, Receiver Operating Characteristic.