The use of deep neural networks is credited with the recent notable improvement in medical image processing accuracy, as manual segmentation is laborious and leads to interpretation errors. For cardiologists, segmenting the atrial is crucial, particularly in the left atrium (LA). Given the likelihood of narrowing in the region where blood flow ceases and a heart attack occurs. The most common cardiac arrhythmia, atrial fibrillation, can be precisely diagnosed and treated with the use of automated approaches that divide the blockage zone.
Although deep convolutional neural networks (DCNNs) have recently demonstrated impressive performance in segmentation tasks, it remains challenging because of how wide the network's receptive field might still be constrained. This restriction may lead to the incorrect classification or insufficient segmentation of items that need more comprehensive contextual data.
In this reseaerch, we propose a novel memory efficient Sparse Attention U-net (SATNet) which can capture long-range dependencies while feasibly reducing the computation cost. In our network, sparse attention blocks allow our encoder-decoder based U-Net model to selectively attend to the relevant information in the skip connections. According to the experimental findings, the sparse attention U-Net performs better than standard segmentation models with a Dice score of 0.93. The outcome shows that the model presented in this paper can successfully raise the standard of left atrial segmentation accuracy, setting the stage for further research into atrial reconstruction.