Self-supervised MRI Image Denoising Method Incorporating Mixed-Attention Based on Residual UNet
There are many different types of noise during MRI imaging,which usually affects the doctor's judgment of the con-dition.Existing deep learning-based MRI image denoising methods require paired images to train the network,which has poor adaptability to noise types.In this paper,a self-supervised MRI image denoising method that incorporated a mixed-attention based residual UNet is proposed.The proposed method first uses a neighbor sampler to obtain two paired noisy images by down-sampling from a single noisy MRI image,then uses the mixed-attention residual UNet network to deeply extract image features.The denoising network is trained by combining the reconstruction loss function and the regularization loss function mentioned in the paper.Tests are deployed in different levels of Gaussian noise and Rician noise levels on Brainweb dataset,respectively.Results show that the denoising ability of this paper's method improves by 4%for Gaussian noise compared with conventional MRI image denoising meth-od.
万方数据
维普
