基于磁共振Dixon图像深度学习进行模拟CT骨重建的研究  

A study of CT skull reconstruction with MR Dixon images based on deep learning method

作　　者：刘克明[1] 曲源[1] 赵洪飞 黄琼 毋晓萌 尚斐[3] LIU Ke-ming;QU Yuan;ZHAO Hong-fei(Department of Radiology,People's Hospital of Xinjiang Uygur Autonomous Region,Urumqi 830001,China)

机构地区：[1]新疆维吾尔自治区人民医院放射影像中心,乌鲁木齐830000 [2]上海交通大学医学院附属瑞金医院,上海200020 [3]北京理工大学生命学院,北京100081

出　　处：《放射学实践》2022年第11期1432-1435,共4页Radiologic Practice

摘　　要：目的:基于磁共振Dixon图像的不同组合,采用深度学习方法进行颅骨二值重建,通过与CT图像比较评估骨重建效果。方法:回顾性收集2021年6月-8月共21例头颅CT和MR图像。刚性配准后,将CT值大于150和400 HU像素点作为颅骨组织。采用U-Net神经网络模型训练,16例作为训练集,5例作为测试集。使用Dixon四种对比图像及其不同组合形成集成模型,进行二值颅骨图像重建。采用戴斯相似性系数(DSC)、准确度、敏感度和特异度评估骨重建效果。结果:在以400 HU为阈值重建MR二值骨图像,水相和同相位组合的重建结果DSC值最高(0.760±0.038)。在以150 HU为阈值时,水相和反相位组合的重建结果DSC值最高(0.795±0.040)。150 HU重建结果比400 HU敏感度高(0.880±0.050 vs.0.855±0.052),特异度下降(0.977±0.004 vs.0.982±0.004)。结论:利用Dixon图像进行深度学习重建颅骨二值图像,在400 HU为阈值时水相和同相位图像结合进行颅骨重建的效果最优,在150 HU为阈值时水相和反相位图像结合的效果最优。Objective:Based on different combinations of MR Dixon images,skull binary reconstruction was performed using deep learning methods,and the effectiveness was evaluated by comparing with CT images.Methods:A total of 21 CT and MR head images were collected retrospectively from June to August 2021.After rigid registration between CT and MR images,pixels with CT values greater than 150 HU and 400 HU were regarded separately as skull tissue.The 2 D U-Net neural network model was applied for skull reconstruction with 16 cases as training set and 5 cases as test set.Four kinds of Dixon images and their different combinations were used to generate an integrated model for training.The results were evaluated by dice similarity coefficient(DSC),accuracy,sensitivity and specificity.Results:Using 400 HU as the threshold for MR binary bone reconstruction,the combination of water image and in phase image showed the highest DSC(0.760±0.038).Using 150 HU as the threshold,the combination of water image and out of phase image showed the highest DSC(0.795±0.040).The results of 150 HU as the threshold were more sensitive than 400 HU(0.880±0.050 vs.0.855±0.052),but the specificity decreased(0.977±0.004 vs.0.982±0.004).Conclusion:MR Dixon images could be used to perform deep learning for the reconstruction of skull binary image.When 400 HU as a skull threshold,the water image and in phase image were combined to show the best skull reconstruction effectiveness.When 150 HU as a threshold,the water image and out of phase image were combined to show the best result.

关 键 词：卷积神经网络 深度学习 磁共振成像 颅骨 图像重建 

分 类 号：R814.42[医药卫生—影像医学与核医学] R445.2[医药卫生—放射医学]