DWT-Net: A Medical Image Segmentation Model Incorporating Frequency Domain Information

Numerous medical image segmentation frameworks have been proposed by researchers; however, the incorporation of frequency domain information has been largely overlooked in previous studies. By incorporating frequency information, we avoid the issue where traditional downsampling reduces computational load and communication bandwidth but also removes redundant and important information, leading to accuracy loss. This paper introduces a novel supervised convolutional neural network architecture, termed “DWT-Net,” which incorporates discrete wavelet convolution and a custom residual downsampling mechanism. Through wavelet transformation, the model not only achieves a near-global receptive field but also significantly enhances its responsiveness to low-frequency information. To bolster the model’s performance, data augmentation techniques were employed to enrich the dataset. Our model has been applied to the public datasets Kvasir-SEG and CVCClinicDB. Compared to other frameworks, it has shown significant improvements in display, achieving the state-of-the-art accuracy.