基于反投影回归网络的医学图像超分辨率重建  

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

作　　者：郑秋富 金燕[1] 姜智伟 胡奕标 Zheng Qiufu;Jin Yan;Jiang Zhiwei;Hu Yibiao(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023)

机构地区：[1]浙江工业大学信息工程学院,杭州310023

出　　处：《计算机辅助设计与图形学学报》2024年第12期1968-1977,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(U22A2040).

摘　　要：针对使用深度学习进行医学图像超分辨率重建时存在的欠定性问题,提出了一种基于反投影回归网络的医学图像超分辨率重建算法.该网络除了学习低分辨率到高分辨率图像的原始映射之外,还学习一个对偶回归映射来预测退化核并重建出低分辨率图像,形成一个闭环以提供额外的约束条件,通过该约束减小低分辨率到高分辨率图像的映射空间,从而缓解图像重建时的欠定性问题;并在上采样和下采样过程中引入了反投影机制,通过误差反馈来减少上、下采样时丢失的特征信息.在PyTorch环境中,与EDSR,DBPN,RCAN等算法在Mayo和TCGA-KICH数据集进行了对比实验.结果表明,所提算法的评价指标优于对比算法,如Mayo数据集上2倍放大因子下,所提算法的峰值信噪比高于对比算法0.12~1.41 dB.To address the ill-posed problem when using deep learning for medical image super-resolution recon-struction,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 upsam-pling 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.

关 键 词：医学图像重建 对偶回归 误差反馈 U-Net 卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]