Medical Image Super-Resolution Reconstruction Based on Back Projection Regression Network

To address the ill-posed problem when using deep learning for medical image super-resolution reconstruction, a medical image super-resolution reconstruction algorithm based on back projection regression network is proposed. In addition to learning the primal mapping from low resolution images to high resolution images, the back projection regression network also learns a dual regression mapping to estimate the degradation kernel and reconstruct low resolution images, forming a closed loop to provide an extra constraint. The constraint reduces the mapping space from low resolution images to high resolution images so as to alleviate the ill-posed problem of super-resolution reconstruction. Besides, the back projection mechanism is introduced in the process of upsampling and downsampling to reduce the lost feature information by error feedback. In the PyTorch environment, the comparison experiments with EDSR, DBPN, RCAN and other algorithms were carried out on Mayo and TCGA-KICH data sets. The results show that the evaluation index of the proposed algorithm is better than the comparison algorithm. For instance, the peak signal to noise ratio of the proposed algorithm is 0.12~1.41 dB higher than the comparison algorithm at 2 times magnification factor on Mayo data set.