Learnable Frequency Filters and Low-Frequency Domain Randomization Based Domain Generalization for Medical Image Segmentation

In recent years, the convolutional neural network represented by U-net has achieved great success in the field of medical image segmentation. However, due to different data acquisition protocols and differences in instrument parameters, the data in actual clinical applications will have a distribution shift problem with the training data of the model, also known as the domain shift problem. This can lead to a decrease in model performance, hindering the application of the model in real-world clinical settings. Traditional convolutional neural networks have local perception properties, and their inability to model long-range dependencies hinders the improvement of domain generalization performance. In addition, the traditional DG method of feature space domain randomization is not effective in medical image segmentation tasks due to the limited search space of feature styles and the inability to keep semantic information unchanged. To address the above issues, this paper proposes a medical image segmentation network named GFDR-U net for domain generalization. The network utilizes globally learnable filters to model long-range dependencies, and introduces a low-frequency domain randomization method to solve the problem of semantic invariance. Extensive experiments show that our method outperforms the state-of-the-art methods.