使用U-Net深度学习网络对腰椎矢状T_(2)WI图像自动分割的可行性研究

U-Net deep learning network for automatic segmentation of lumbar structures on sagittal T_(2)WI images

目的:研究训练U-Net模型自动分割腰椎矢状面T_(2)WI图像中各结构的可行性。方法:回顾性搜集腰椎矢状面T2WI图像数据,共获得80个矢状面T_(2)WI序列。由2位影像医师手工标注矢状面腰椎椎体、椎间盘、椎间孔、椎管/硬膜囊、脊髓及马尾神经。将数据随机分为训练集、调优集和测试集,使用U-Net网络分两步(coarse-to-fine)训练腰椎矢状T_(2)WI分割模型。模型评价指标包括客观评估(Dice系数)和主观评估。结果:11例测试集数据中U-Net模型预测腰椎5个解剖部位分割的Dice值分别为椎体0.82~0.9(平均0.864)、椎间盘0.86~0.92(平均0.898)、椎管/硬膜囊0.76~0.87(平均0.837)、椎间孔0.6~0.76(平均0.67)、脊髓及马尾神经0.55~0.9(平均0.669)。主观评估各解剖部位分割满意率分别为椎体97.5%、椎间盘97.9%、椎管/硬膜囊86.4%、椎间孔76.7%、脊髓及马尾神经78.6%。结论:基于U-Net深度学习网络对腰椎矢状T_(2)WI图像的解剖结构进行自动分割是可行的。

Objective:To investigative the feasibility of training a U-Net model for automatic segmentation of lumbar spine structures on sagittal T_(2)WI MR images.Methods:The sagittal MRI images were retrospectively collected.A total of 80 sagittal T_(2)WI sequences were obtained.The lumbar spine column,intervertebral disk,intervertebral foramen,lumbar spinal canal/subarachnoid space,spinal cord and cauda equina were manually annotated by 2 radiologists on sagittal T_(2)WI images.The data were randomly divided into train set,validate set,and test set.A cascade U-Net network was used to develop the segmentation model(coarse-to-fine)of the lumbar spine structures.The evaluation indexes of the model include objective evaluation(dice coefficient)and subjective evaluation.Results:In the test set of 11 cases,the dice values of the U-Net model for the 5 lumbar spine structures were as follows:spine column 0.82~0.9(mean value 0.864),intervertebral disk 0.86~0.92(mean value 0.898),lumbar spinal canal/subarachnoid space 0.76~0.87(mean value 0.837),intervertebral foramen 0.6~0.76(mean value 0.67),spinal cord and cauda equina 0.55~0.9(mean value 0.669).Subjectively,the satisfaction rates of lumbar spine structures'segmentation were as follows:spine column 97.5%,intervertebral disk 97.9%,lumbar spinal canal/subarachnoid space 86.4%,intervertebral foramen 76.7%,spinal cord and cauda equine 76.7%.Conclusion:It is feasible to segment the lumbar structures of sagittal T_(2)WI image automatically based on U-Net deep learning network.