残差网络-迁移学习法抑制低剂量CT噪声  

Low-dose CT Noise Suppression with Residual network-Transfer learning

作　　者：钟安妮 李斌 罗宁[1] 徐圆[1] 周凌宏[1] 甄鑫[1] ZHONG An-ni;LI Bin;LUO Ning;XU Yuan;ZHOU Ling-hong;ZHEN Xin(School of Biomedical Engineering,Southern Medical University,Guangzhou 510515,China;Department of Radiation Oncology,Sun Yat-sen University Cancer Center,State Key Laboratory of Oncology in South China,Collaborative Innovation Center for Cancer Medicine,Guangzhou 510060,China)

机构地区：[1]南方医科大学生物医学工程医疗仪器研究所,广州510515 [2]中山大学肿瘤防治中心,华南肿瘤学国家重点实验室,肿瘤医学协同创新中心,广州510060

出　　处：《核电子学与探测技术》2019年第4期452-456,共5页Nuclear Electronics & Detection Technology

基　　金：国家自然科学基金(81728016、81874216、81601493、81571771);国家重点研发计划(2017YFC 0112903)资助

摘　　要：针对低剂量CT图像的噪声伪影问题,提出一种残差网络-迁移学习的图像噪声抑制算法。采用残差卷积神经网络框架,学习从低剂量CT图像到其噪声图像的映射关系。用自然图像对该模型进行简单的预训练;将CT图像数据对模型的共享参数进行微调;从低剂量CT图像中减去所得到的噪声图像,实现噪声抑制。实验表明:该方法能快速有效地将自然图像域的噪声抑制能力迁移到医学图像域,量化指标均优于传统的去噪算法,可以有效抑制噪声和保留图像的细节结构。To remove the noise from LDCT,a residual network with transfer learning is proposed.In our framework,the residual convolutional neural network is utilized,aiming at learning the matching from LDCT to noise map.The model is firstly trained by noisy natural images,then fine-tuned by LDCT training set.Finally,the noise is subtracted from the low dose CT image to achieve artifact suppression.The experiments confirm that the proposed network can transfer the noise suppression ability of natural image domain to medical image quickly and effectively.And the quantitative index outperforms traditional denoising algorithms,with better artifact suppression as well as the image detail preservation.

关 键 词：低剂量CT CT图像去噪 迁移学习 残差网络 

分 类 号：TN911.73[电子电信—通信与信息系统]