医学图像的深度先验重建仿真  

Deep Prior Reconstruction Simulation of Medical Images

作　　者：李明杰 吕伟 易正明[1] LI Mingjie;LÜWei;YI Zhengming(State Key Laboratory of Refractories and Metallurgy,Key Laboratory for Ferrous Metallurgy and Resources Utilization of Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,China)

机构地区：[1]武汉科技大学耐火材料和冶金国家重点实验室,钢铁冶金与资源利用教育部重点实验室,武汉430081

出　　处：《实验室研究与探索》2023年第3期40-45,共6页Research and Exploration In Laboratory

摘　　要：深度图像先验算法(DIP)可以将代数迭代过程与深度神经网络有效结合,完成高质量的图像重建。构建了一种基于“Conformer”结构的DIP重建算法,该网络结构改善了原始DIP算法在重建中出现的伪影问题,同时提高了算法对于病理图像在全局特征和局部特征中的重建能力。提出了基于感知特征相似性度量的正则化方法,使重建图像与原始图像之间的特征差异达到最小,提高图像特征细节的还原度。经对比发现,该算法在低剂量平行束CT数据集上将峰值信噪比提高了2~5 dB,在提高图像内部纹理的同时保留了良好的边缘细节。Deep image prior(DIP)algorithm can effectively combine algebraic iterative process by deep neural network,so as to complete high-quality image reconstruction.A DIP reconstruction algorithm based on“conformer”structure is constructed.This network structure improves the artifact problem of the original DIP algorithm in the reconstruction,and improves the ability of the pathological image reconstruction in the global and local features.A regularization method based on perceptual feature similarity measure is proposed.Its function is to minimize the feature difference between the reconstructed image and the original image and improve the restoration degree of image feature details.By comparison,it is found that the algorithm improves the peak signal-to-noise ratio(PSNR)by 2-5 d B in low-dose parallel beam CT datasets,and preserves edge details well while improving the internal texture of the image.

关 键 词：断层扫描成像 重建算法 深度图像先验 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]