Research On Medical Image Super-Resolution Reconstruction Based On Residual Network

Medical image is import reference information and widely used in the clinical.The better treatment decision will be made by doctor because of the help from high resolution medical image.However,it is difficult to obtain high resolution medical image in the context of using low precision medical imaging equipment.The high resolution medical image can be generated in a low-cost way from the low resolution one used super-resolution reconstruction technology.The medical task requires high accuracy.The feature of textural in medical image can be reconstructed better by residual network.It will focus on the problem of medical image super-resolution reconstruction used residual network structure in the dissertation.The main research content as follows:(1)The new network structure aimed in medical image super-resolution reconstruction base on residual dense has been proposed.It improves the skip connection type in the original residual dense blocks.Two different skip connection types are established for transmitting features map: close skip connection and remote one.The close one batch two convolution layers and established connection.The remote one established connection by double geometric series.At the same time the activation function is substituted.A simple ablation experiment is designed for comparing the contribution of different methods for improvement.The complexity of connection declines to logarithm from exponent.The problem that the proportion of shallow features is too high has been resolved.The results of experiment show that the average PSNR and SSIM is improved by 0.97 d B and 0.0219 respectively in condition of comparing with residual dense and x3 scale factor.(2)A medical image super-resolution reconstruction network model base on residual network and meta-learning is introduced.With the previous work,a new structure of features map unsampled that using the strategy of meta-learning is proposed.The model can be train together with a simple neural network embedded by the coordinate and scale information.The operation space of fractional and integral scale tasks is unified by vector duplication.The problem that a model only matches one integral scale and do not support the fractional scale has been resolved.The experiments including baseline and ablation are designed and conducted in three different medical images.The results of experiment show that the average PSNR and SSIM is improved by 1.87 d B and 0.0466 respectively in condition of comparing with sub-pixel and test set b for x3 scale factor.