CAW-CL: Cascaded Adaptive Weighted Contrastive Learning for Unsupervised Ultrasound Plane-Wave Image Reconstruction

Ultrasound plane-wave (PW) imaging achieves an ultra-high frame rate at the expense of degraded image quality. Deep learning-based methods have emerged as a promising avenue for enhancing PW imaging quality. Existing research predominantly focuses on supervised learning that relies on high-quality paired references. This is challenging in real-world scenarios, highlighting the need for unsupervised learning techniques to reconstruct high-quality images with unpaired data. Considering the physical characteristics of ultrasound signals, where image degradation varies across different echo intensities, existing unsupervised approaches struggle to differentiate this variability. In this paper, we propose a novel unsupervised method based on Cascaded Adaptive Weighted Contrastive Learning (CAW-CL) to deal with signal degradation associated with different intensity echo areas. To achieve consistent improvement across different intensity echo areas, we propose the adaptive weight matrix network (AWMNet) with Swin-Transformer blocks for dynamic attention to patches with varying sizes and positions. Furthermore, a cascaded structure helps AWMNets integrate information across feature layers more effectively. Experimental results demonstrate that our model outperforms several learning-based methods in both qualitative and quantitative evaluations and shows robust reconstruction ability in diverse intensity echoes.