摘要
目的:探讨3D U-Net模型自动分割颈椎矢状面T_(1)WI和T_(2)WI图像上颈椎各结构的可行性。方法:回顾性搜集拟诊为颈椎病的92例患者的矢状面T_(1)WI和T_(2)WI图像资料,由两位影像医师在每例患者的2个序列图像上分别人工标注颈椎各结构,包括椎体、椎间盘、硬膜囊、脊髓和椎间孔。将178个序列的图像随机分为训练集(n=138)、调优集(n=20)和测试集(n=20)。采用训练集的数据训练3D U-Net分割模型,在调优数据集中微调参数,在测试集中采用定量指标(Dice相似系数,DSC)和定性指标(主观评分)评价模型的分割效能,并比较各结构的DSC值在3组内及3组间是否存在统计学差异。结果:在测试集中3D U-Net模型分割颈椎椎体、椎间盘、硬膜囊、脊髓和椎间孔的DSC值分别为0.87±0.03、0.85±0.04、0.87±0.04、0.82±0.05和0.57±0.08,分割颈椎各解剖结构的总体DSC值为0.80±0.13。各结构的DSC值在3组内及组间均有统计学差异(P<0.001)。主观评价显示3D U-Net模型分割颈椎各结构获得的图像均符合临床测量要求。结论:基于矢状面T_(1)WI和T_(2)WI序列的3D U-Net模型对颈椎各结构的分割可达到较高的准确性。
Objective:To investigative the feasibility of 3D U-Net model for automatic segmentation of anatomic structures of cervical spine on sagittal T_(1)WI and T_(2)WI MR images.Methods:The sagittal T_(1)WI and T_(2)WI images of 92 patients(178 series)were retrospectively collected.The cervical spine column,intervertebral disc,subarachnoid space,spinal cord and intervertebral foramen were manually drawn and annotated by 2 radiologists on sagittal T_(2)WI and T_(1)WI images.The images were randomly assigned to the dataset of training(n=138),validate(n=20)and test(n=20).In the trai-ning dataset,a 3D U-Net model was trained with the training dataset.The model was fine-tuned with the validation dataset.In the test dataset,the segmentation efficacy of the model was evaluated by quantitative index(Dice similarity coefficient,DSC)and qualitative evaluation.The DSC value of each structure were statistically compared within and among the three datasets.Results:In the test set,the DSC values of the trained 3D U-Net model in segmentation of the cervical vertebral body,intervertebral disc,subarachnoid space,spinal cord and intervertebral foramen were 0.87±0.03,0.85±0.04,0.87±0.04,0.82±0.05,and 0.57±0.08,respectively.The general DSC value in segmentation of the cervical vertebral structures was 0.80±0.13.The DSC value of each structure was statistically different within and among the three datasets(all P<0.001).The qualitative scores met the needs of clinical application.Conclusion:The 3D U-Net model based on sagittal T_(1)WI and T_(2)WI MR images can achieve good performance for automatic segmentation of anatomic structures in cervical spine.
作者
朱逸峰
赵凯
郭丽
张耀峰
王祥鹏
张晓东
李雨师
王霄英
ZHU Yi-feng;ZHAO Kai;GUO Li(Medical Imaging Center,the Peking University First Hospital,Beijing 100034,China)
出处
《放射学实践》
CSCD
北大核心
2021年第12期1558-1562,共5页
Radiologic Practice
关键词
磁共振成像
颈椎
深度学习
人工智能
自动分割
Magnetic resonance imaging
Cervical vertebral
Deep learning
Artificial intelligence
Automatic segmentation
