CAMDA Net: Cross Attention Multi-modal Domain Adaptation using B-mode and Sinogram

Lung ultrasound (LUS) is a valuable imaging technique in point-of-care (POC) for the prompt diagnosis of lung disease, but accurately identifying specific diseases can be challenging. Recently, learning-based lung disease classification networks have shown potential for clinical applications. However, current approaches rely on supervised learning using LUS data from cart-type devices, leading to a domain gap when using lower-quality LUS from POC hand-held (HH) devices. In this paper, we proposed Cross Attention Multi-modal Domain Adaptation (CAMDA) lung classification network, which aims to address this domain gap. CAMDA-net leverages both B-mode and sinogram modalities, using cross-attention to focus on domain-invariant biomarkers while minimizing device-specific dependencies. The proposed domain adaptation scheme ensures robust performance even with unseen LUS data from HH devices. Quantitative results demonstrate that CAMDA-net significantly enhances balanced accuracy by 35.4% over baseline models, confirming its effectiveness in mitigating domain discrepancies.