自监督深度学习的心脏磁共振图像配准算法  

Self-Supervised Deep-Learning Cardiac Magnetic Resonance Image Registration Algorithm

作　　者：刘子兴 廉钰 李汉军[1] 唐晓英[2] LIU Zixing;LIAN Yu;LI Hanjun;TANG Xiaoying(School of Life Science,Beijing Institute of Technology,Beijing 100081,China;School of Medical Technology,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]北京理工大学生命学院,北京100081 [2]北京理工大学医学技术学院,北京100081

出　　处：《中国医疗设备》2024年第11期27-32,38,共7页China Medical Devices

摘　　要：目的通过使用合成图像的方法解决在配准过程中缺少金标准的问题,并应用深度学习算法进行心脏T_(1)定量图配准。方法首先利用T_(1)加权图像的先验信息合成无运动的参考图像;其次使用DeepIPMCNet卷积神经网络来学习和配准层内运动。另一个网络DeepTPMDNet用于检测和消除穿层运动。使用在自由呼吸条件下采集的STONE序列T_(1)映射数据集进行训练、验证和测试,以验证本文方法的有效性。通过T_(1)标准差和SD map标准差来评估性能。结果在配准后,左心室和室间隔的Dice系数、T_(1)标准差和SD map标准差均得到了改善(通过DeepIPMCNet,左心室的Dice系数从0.88提高到0.90,室间隔的T_(1)标准差从121.91 ms降低到86.99 ms,SD map标准差从46.49 ms降低到36.53 ms;通过DeepTPMCNet,左心室的Dice系数从0.74提高到0.93,室间隔的T_(1)标准差从192.02 ms降低到114.37 ms,SD map标准差从93.41 ms降低到50.53 ms),差异均有统计学意义(P<0.001)。结论本研究提出的深度学习方法可有效缓解心脏和呼吸运动对心脏T_(1)定量图的影响。Objective To address the lack of a gold standard in registration by employing synthetic images,and to perform the the registration of cardiac T_(1) map by deep learning algorithms.Methods Firstly,the motion-free reference images was synthesized by using prior information from T_(1)-weighted images.Subsequently,a DeepIPMCNet convolutional neural network was utilized to learn and register in-plane motion.Another network DeepTPMDNet was employed for the detection and elimination of through-plane motion.The STONE sequence T_(1) mapping dataset collected under free-breathing conditions was used for training,validation,and testing,so as to validate the effectiveness of the proposed method.Performance was evaluated by T_(1) standard deviation and SD map standard deviation.Results After registration,the Dice coefficient,T_(1) standard deviation and SD map standard deviation of left ventricle and ventricular septum were improved(Using DeepIPMCNet,Dice coefficient of the left ventricle increased from 0.88 to 0.90,T_(1) standard deviation of the interventricular septum decreased from 121.91 ms to 86.99 ms,and SD map standard deviation decreased from 46.49 ms to 36.53 ms;Using DeepTPMCNet,Dice coefficient of the left ventricle increased from 0.74 to 0.93,T_(1) standard deviation of the interventricular septum decreased from 192.02 ms to 114.37 ms,and SD map standard deviation decreased from 93.41 ms to 50.53 ms).These differences were statistically significant(P<0.001).Conclusion The deep learning approach proposed in this study effectively alleviates the effects of cardiac and respiratory motion on cardiac T_(1) quantitative mapping images.

关 键 词：心脏磁共振(CMR) T_(1)定量图 配准算法 自监督深度学习 卷积神经网络 DeepIPMCNet DeepTPMDNet 

分 类 号：R197.3[医药卫生—卫生事业管理] TP391[医药卫生—公共卫生与预防医学]