CTUnet: A Novel Paradigm Integrating CNNs and Transformers for Medical Image Segmentation
Xiaoyu Wang, Chunlin Zhu, Jiaquan Li · 2024
In recent years, convolutional neural networks (CNNs) have demonstrated remarkable performance in medical image segmentation tasks, particularly those utilizing the U-shaped neural network architecture. However, the inherent limitations of CNNs, such as the inability to establish long-range dependencies and the loss of spatial information during downsampling, have constrained their progress in medical image segmentation. Subsequently, researchers have achieved significant success by replacing the convolutional operations in U-Net with Transformers, which can establish long-range dependencies. But, for medical image segmentation tasks, simultaneously modeling global relationships and extracting local details is crucial. Therefore, in this paper, we propose a novel CTBlock that combines CNN and Transformer as the foundational encoder. This method is capable of providing different levels of semantic information and local features while mitigating the loss of spatial information in downsampling. Additionally, we have developed a functional module called CTFusion, which integrates local features extracted from CNN and long-range dependencies from Transformer. It replaces skip connections in U-Net. Subsequently, based on the U-Net architecture, we utilized CTBlock and CTFusion functional modules to construct CTUnet. Finally, we validated our model on the BUSI dataset and DDTI medical dataset, achieving state-of-the-art performance. The code is available at: https://github.com/WXY-Belief/CTUnet.