Related Experiment Videos
Learning generalizable representations across Heterogeneous Acquisition Environments for Breast Ultrasound Diagnosis
Huanjun Wang1, Shukang Zhang1, Wentao Kong2
1College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, Key Laboratory of Brain-Machine Intelligence Technology, Ministry of Education, Nanjing, Jiangsu, 211106, China.
Summary
This study introduces a new framework for breast ultrasound image classification. It improves diagnostic accuracy and consistency by separating image variations from actual pathology, leading to more reliable results.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Machine Learning
Background:
- Ultrasound image classification faces challenges due to variations in image acquisition.
- These variations can lead to inconsistent model performance and limited generalizability in clinical practice.
Purpose of the Study:
- To develop a representation learning framework to disentangle pathology-relevant features from acquisition-style bias.
- To improve diagnostic consistency across heterogeneous breast ultrasound acquisition conditions.
Main Methods:
- Proposed HAE-BUS (Heterogeneous Acquisition Environments for Breast Ultrasound) framework.
- Utilized Multimodal Large Language Models (MLLMs) as an offline semantic interface.
- Inferred latent acquisition-style environments to suppress acquisition-style shortcuts during training.
Main Results:
- Achieved higher Area Under the Curve (AUC) and accuracy compared to existing methods.
- Demonstrated more consistent performance across heterogeneous acquisition conditions.
- Indicated improved robustness and potential for reliable clinical translation.
Conclusions:
- The HAE-BUS framework effectively addresses challenges in breast ultrasound image classification.
- The approach enhances model robustness and diagnostic consistency.
- This method shows promise for improving the clinical translation of AI in medical imaging.
