Knowledge-Aware Multi-Source Domain Adaptation Network for Osteoporosis Diagnosis from Ultrasonic Radio Frequency Signals

Ultrasonic radio frequency (RF) signal can be utilized in the diagnosis of osteoporosis (OP), thereby assisting healthcare professionals in enhancing the screening rates. However, the heterogeneity of the data collected from different centers can significantly affect the accuracy performance of the model, resulting in lower generalization performance in the target domain. We propose a knowledge-aware unsupervised domain adaptation network guided by knowledge perception that considers both common features and domain-specific features. We employ maximum mean discrepancy (MMD) loss to mitigate distribution discrepancies between various domains. In addition, major clinical risk factors (MCRF) are embedded in domain-specific feature extractors to better extract domain-invariant features. We collect in-house datasets from three distinct centers. Our model demonstrates promising results in both single-source and multi-source domain scenarios.