Does TF-Unet improve segmentation accuracy of cardiac MRI images compared to Swin-Unet?
TF-Unet, combining U-Net and Transformer architectures, improves automated cardiac MRI segmentation accuracy compared to Swin-Unet.
Personalized heart models are widely used to study the mechanisms of cardiac arrhythmias and have been used to guide clinical ablation of different types of arrhythmias in recent years. MRI images are now mostly used for model building. In cardiac modeling studies, the degree of segmentation of the heart image determines the success of subsequent 3D reconstructions. Therefore, a fully automated segmentation is needed. In this paper, we combine U-Net and Transformer as an alternative approach to perform powerful and fully automated segmentation of medical images. On the one hand, we use convolutional neural networks for feature extraction and spatial encoding of inputs to fully exploit the advantages of convolution in detail grasping; on the other hand, we use Transformer to add remote dependencies to high-level features and model features at different scales to fully exploit the advantages of Transformer. The results show that, the average dice coefficients for ACDC and Synapse datasets are 91.72 and 85.46%, respectively, and compared with Swin-Unet, the segmentation accuracy are improved by 1.72% for ACDC dataset and 6.33% for Synapse dataset.
Fu et al. (2022) studied this question.