MCTE: MARRYING CONVOLUTION AND TRANSFORMER EFFICIENTLY FOR END-TO-END MEDICAL IMAGE SEGMENTATION

The purpose of the medical image segmentation task is to delineate different organs or lesion regions in the image, which is an important aid for intelligent clinical medical diagnosis. Recent approaches suffer from the inability to obtain reliable attention, are computationally intensive, and do not exploit the relationships between different samples. We marry convolution and Transformer effectively to establish MCTE for medical image segmentation. The proposed MCTE is an end-to-end network based on U-Net with a parallel learning of three types of attention, namely local attention learning with channel and spatial dimensional convolution, global attention learning with smaller computational effort of swin transformer, and external attention learning with two shared memory storing all medical image information. Extensive experimental results on the ACDC and Synapse dataset, which are widely used for the evaluation of medical image segmentation methods, demonstrate that our proposed method exceeds the compared baseline.