Improving the clarity of medical images is of great significance for doctors to quickly diagnose and analyze the disease. However, the existing image denoising algorithms largely depend on the size of the data set, the optimization effect of the loss function, and the difficulty in
adjusting the parameters. Therefore, a medical image denoising algorithm based on deep learning image quality evaluation is proposed. First, the convolution layer of the convolutional neural network and the output of the first full connection layer are used as the perception features. By stacking
the perception loss and pixel loss, and multiplying the perception loss by a certain weight, the low and high level loss fusion of the denoising network is realized, so that the restored image is more in line with human perception. Secondly, by introducing empty convolution into the denoising
network, the mixed expanded convolution kernel and the ordinary convolution kernel are used together in the first layer to increase the range of sensing field. Then, the feature extraction and the quality score regression are integrated into the same optimization process. Finally, the direct
training reconstructed image is transformed into a training noise filter, which reduces the training difficulty and speeds up the convergence of network parameters. The experimental results show that the PSNR and SSIM of the proposed method are 31.63 db and 89.15%, respectively. Compared with
other new image denoising methods, the proposed method can achieve better denoising effect.
Arterial spin labeling MR image denoising and reconstruction using unsupervised deep learning