Learning generalizable representations across Heterogeneous Acquisition Environments for Breast Ultrasound Diagnosis

Ultrasound image classification remains challenging in clinical practice, given substantial variations in image appearance across devices, acquisition protocols, and operator habits, often unrelated to underlying pathology and associated with inconsistent model performance and limited generalizability. To address this challenge, we propose HAE-BUS (Heterogeneous Acquisition Environments for Breast Ultrasound), a representation learning framework that aims to disentangle pathology-relevant features from acquisition-style bias and improve diagnostic consistency across heterogeneous acquisition conditions. Specifically, HAE-BUS uses Multimodal Large Language Models (MLLMs) as an offline semantic interface to characterize pathology-excluded acquisition cues, from which latent acquisition-style environments are inferred without relying on metadata. These inferred environments are then composed during training to suppress acquisition-style shortcuts while preserving pathology-relevant diagnostic features. Experiments conducted on two breast ultrasound benchmarks, including the public BUSBRA dataset and our private NDTH dataset, indicate that our approach obtains higher AUC and accuracy than the compared methods in the evaluated settings, along with more consistent performance across heterogeneous acquisition conditions, indicating improved robustness and suggesting potential for more reliable clinical translation.