基于深度学习模型的颈椎MR图像脊髓及椎管自动测量  

Automatic measurement of spinal cord and canal on cervical MRI images based on deep learning models

作　　者：马超 冯小晨 杨家诚 林丽娜 张雪媛 王晓雯 任菲[4] 邵成伟 曹鹏 曹凯 MA Chao;FENG Xiao-chen;YANG Jia-cheng(Department of Diagnostic Radiology,the First Affiliated Hospital of PLA Naval Medical University,Shanghai 200433,China)

机构地区：[1]中国人民解放军海军军医大学第一附属医院放射诊断科,上海200433 [2]中国人民解放军海军军医大学第二附属医院脊柱外科,上海200003 [3]重庆知见生命科技有限公司,重庆401329 [4]中国科学院计算技术研究所处理器芯片全国重点实验室,北京100190

出　　处：《放射学实践》2024年第11期1514-1520,共7页Radiologic Practice

基　　金：国家自然科学基金面上项目(82372045);上海市自然科学基金面上项目(23ZR1478400)。

摘　　要：目的:探讨深度学习实现颈椎MR脊髓和椎管自动测量的可行性。方法:回顾性收集558例颈椎MR图像,以8:1:1的比例随机分为训练集(n=436)、调优集(n=61)和测试集(n=61)。由一位低年资医师标注所有图像的椎管和脊髓,在测试集中测量脊髓最大受压程度、椎管最大狭窄程度、横截面积及压缩比,由一位主任医师审核所有结果后作为金标准。另一位高年资医师在测试集中进行测量作为人工组结果。以Swin Transformer为骨干网络的深度学习模型进行的分割和测量作为模型组结果。采用Dice相似系数(DSC)、交并比(IoU)评价模型分割性能。使用组内相关系数(ICC)、Bland-Altman散点图比较各组结果一致性。结果:测试集中,深度学习模型分割脊髓(横轴面、矢状面)和椎管(矢状面)的DSC值(%)为93.10±0.57、94.60±0.09、86.17±0.22,IoU值(%)为87.09±1.00、89.76±0.17、75.70±0.34。人工组、模型组和金标准的ICC值为0.770~0.945,模型组与金标准的组间ICC值为0.782~0.913,人工组与金标准的组间ICC值为0.692~0.903,三组之间的差异具有统计学意义(P<0.001)。结论:深度学习模型对颈椎MR椎管和脊髓的分割测量具有较高准确性。Objective:To explore the feasibility of automatic measurement of spinal cord and canal in cervical MR using deep learning.Methods:558 cases of cervical spine MR images were retrospectively collected and randomly divided into training set(n=436),validation set(n=61),and test set(n=61)at 8:1:1.The spinal canal and spinal cord of all images were labeled by a junior physician,and the maximum spinal cord compression,maximum canal compromise,transverse area,and compression ratio were measured in the test set.All results were then reviewed by a chief physician as the gold standard.Another senior physician conducted measurements on the test set as the manual group results.A deep learning model,utilizing Swin Transformer as the backbone network,was employed for segmentation and measurement,contributing to the model group results.Evaluation of the model's segmentation performance was conducted using dice similarity coefficient(DSC)and Intersection-over-Union(IoU).Consistency was compared across groups,including the junior physician,chief physician,manual group,and the deep learning model,using Intraclass correlation coefficient(ICC)and Bland-Altman scatter plots.Results:In the test set,the DSC values(%)for the deep learning model segmentation of the spinal cord(axial and sagittal planes)and vertebral canal(sagittal plane)were 93.10±0.57,94.60±0.09,and 86.17±0.22,respectively.The IoU values(%)were 87.09±1.00,89.76±0.17,and 75.70±0.34,respectively.The ICC values of the manual group,model group,and gold standard ranged from 0.770 to 0.945.The ICC values of the model group and gold standard ranged from 0.782 to 0.913;the ICC values of the manual group and gold standard ranged from 0.692 to 0.903.The differences between the three groups were statistically significant(P<0.001).Conclusion:The deep learning model demonstrates high accuracy in the segmentation and measurement of spinal cord and canal in cervical MR.

关 键 词：磁共振成像 颈椎 深度学习 自动分割 自动测量 

分 类 号：R445.2[医药卫生—影像医学与核医学] R-05[医药卫生—诊断学]