Deep Residual Network Based Medical Image Reconstruction

Recent years have witnessed the development of medical imaging technology. However, the process of imaging, storage and transmission often makes the image quality reduced and affects the visual and post-processing effect. The degradation of medical image often leads to the interference of noise and the decrease of resolution. In order to reconstruct the degraded medical image, a deep residual network combined with perceptual loss and mean square error (MSE) loss is proposed to enhance image quality. As a result, a single network can handle de-noising and super-resolution in same time. By using the residual network, the number of network layers can be deepen while the gradient dispersion problem can be avoided. More image edges and details can be reconstructed with the joint loss. Experiments on one medical image data set TCIA show that the proposed method can jointly perform de-noising and super-resolution to restore more medical image texture details and get better visual effect, especially for restraining noise in the low-dose CT image.