Local Spatial Attention Transformer For Sparse Photoacoustic Image Reconstruction

Sparse sampling is crucial in photoacoustic imaging for rapid data acquisition with low computation. However, reconstructing high-quality images from sparse data is challenging due to undersampling artifacts. Traditional convolutional neural networks, while useful in image reconstruction, show limitations in capturing the global contextual information necessary for effective photoacoustic tomography (PAT) image reconstruction. Here, we propose a novel solution that employs the neighborhood attention (NA) mechanism for sparse PAT reconstruction. This approach highlights the importance of local neighborhood information and spatial context in reconstructing sparse photoacoustic data. Unlike channel attentionbased methods, our spatially-based NA transformer, designed with a sliding kernel window, can recover parts of an image, such as the boundary, even in extremely sparse sampling scenarios. Our model, built on the NA mechanism, amplifies relevant features and suppresses undersampling artifacts, facilitating robust and high-quality image reconstructions. This work marks the first application of the sliding window attention mechanism to image restoration tasks in PAT. This research highlights the potential of spatial attention mechanisms, particularly the NA mechanism, in sparse reconstruction for PAT, paving the way for potential enhancements in other imaging modalities