Medical Image Denoising Algorithm Via Deep Learning

With the rapid development of modern diagnostic medical technology,medical imaging technologies such as MRI and electronic computed tomography have become important tools for medical diagnosis.MRI is an important basis for assisting clinical medical treatment by obtaining soft tissue images and detecting lesions in vivo through in vitro imaging.However,MRI is affected by uncontrollable factors such as detection equipment and human activities,and is accompanied by image quality degradation problems such as noise and artifacts,which affect disease diagnosis.A medical image denoising method based on deep learning is proposed for the quality degradation problems such as noise and artifacts of medical images,and the main research contents are as follows:(1)For the noise and checkerboard artifact interference problems,this paper proposes an improved U-Net network architecture.First,the downsampling module in the improved U-Net network architecture uses patch merging instead of maximum pooling to solve the image information loss problem.Second,the upsampling module in the U-Net network architecture is improved by introducing dual upsampling(bilinear sampling and subpixel convolution)to avoid the tessellation artifact problem caused by the block effect of transposed convolution.Finally,the network structure is deepened using residual connections to improve the expressiveness of the network while obtaining contextual and location information to solve the lesion localization problem.(2)To address the problems of high computational complexity of high-resolution feature maps and blurred edges of denoised images,this paper proposes a Swin-Conv-UNet(SCU)method,which combines Swin-Transformer with powerful global modeling capability and convolutional neural network with capturing local features,and embeds them in the above improved Unet architecture.First,a shallow feature extraction module is constructed to obtain low-frequency information such as color or texture of the input image by a single convolutional layer.Secondly,the Unet feature extraction module is constructed to feed the shallow features into Unet to extract the high level and multi-scale deep features to prevent the loss of image detail information.Finally,the reconstruction module is constructed to perform high-quality image reconstruction by aggregating shallow and deep features.The experimental results show that the improved deep learning medical image denoising model based on Swin-Transformer has high denoising accuracy and fast running speed,which can meet the medical image denoising tasks in many scenarios and provide strong technical support and services for doctors’ medical diagnosis.