AgentUNet:A U-shaped Hybrid Network with CNN-Transformer for Medical Image Segmentation

An accurate and efficient medical image segmentation algorithm plays an important role when doctors diagnose diagnosis. Using deep learning methods to tackle medical image segmentation tasks holds great significance for both medicine and computer science. UNet is one of the most important achievements in the field of medical image segmentation using deep learning. However, merely relying on the U-shaped architecture of UNet can struggle to meet clinical requirements for processing fine details and small lesions in information-rich medical images. To address this limitation, we propose a novel segmentation architecture, AgentUNet, which enhances the global feature extraction module and introduces an innovative SCD module for integrating local and global features. It can effectively capture the edge details of objects of varying sizes and fuses semantic information from both the encoder and decoder, thereby minimizing the potential loss of high-level semantic information. We evaluated our method on two publicly available datasets, Synapse and ACDC, and achieved superior results compared to other SOTA models.