TELA-UNet: U-Net Augmented with Transformer and Efficient Local Attention for Medical Image Segmentation

With the continuous advancement of deep learning technology in medical image processing, precise image segmentation techniques have become increasingly important. Currently, TransUNet, which combines CNN and Transformer, performs well in automatic segmentation tasks. However, TransUnet lacks the ability to utilize spatial features within images. To address these issues, this paper introduces a new medical image segmentation framework, TELA-UNet, designed to integrate efficient local attention blocks (ELA-Blocks) with TransUNet to enhance segmentation accuracy. Specifically, by incorporating ELA-Blocks in both the embedding layer and each skip connection, the feature extraction capability is significantly enhanced, thereby improving the efficiency of the encoder-decoder structure. Experimental results demonstrate that the proposed method achieves a Dice Similarity Coefficient (DSC) of 80.35% and an average Hausdorff Distance (HD) of 21.41mm on the Synapse dataset, surpassing several state-of-the-art segmentation methods.