基于腰椎螺旋CT图像以卷积神经网络技术全自动识别并重建椎间盘的可行性

目的观察基于腰椎螺旋CT图像以卷积神经网络技术全自动识别及重建椎间盘的可行性。方法回顾性分析400例腰痛患者的腰椎CT资料,以其中320例为训练集、40例为验证集、40例为测试集。以人工智能(AI)系统进行学习训练和测试。以深度学习(DL)卷积神经网络3D V-Net技术分割腰椎轴位CT图像中的椎体与椎间盘,并轴位重建椎间盘;以Dice系数评估分割精度。由2名放射科医师分别对AI重建图像及人工重建图像进行图像质量评分并进行对比。结果AI分割骶椎椎体、L5椎体、L1~L4椎体及椎间盘的Dice系数分别为0.953、0.940、0.940及0.926,平均为0.940。针对测试集40例,采用腰椎螺旋CT数据经卷积神经网络技术完成197个椎间盘重建。2名放射科医师对197幅AI重建图像及人工重建图像的中位评分均为4分,差异无统计学意义(P均>0.05);评分一致性加权Kappa值为0.86295%CI(0.778,0.946),P<0.001]。结论基于腰椎螺旋CT图像卷积神经网络全自动识别及重建椎间盘的可行性令人满意。

Feasibility of automatic recognition and reconstruction of intervertebral disc based on lumbar spiral CT images using convolutional neural network

Objective To observe the feasibility of automatic recognition and reconstruction of intervertebral disc based on lumbar spiral CT images using convolutional neural network. Methods Data of lumbar spiral CT of 400 patients with low back pain were retrospectively analyzed. Then 320 cases were included in the training set, 40 cases in the validation set and 40 cases in the test set. The artificial intelligence (AI) system was used for data training and testing. A deep learning (DL) convolutional neural network 3D V-Net technology was used to segment vertebral body and intervertebral disc on axial CT images of lumbar spine, then intervertebral discs were reconstructed at axial position. Dice coefficient was used to evaluate the accuracy of segmentation. The quality of AI reconstructed and manually reconstructed images were rated were compared by 2 radiologists, respectively. Results Dice coefficient of the sacral vertebral body, L5 lumbar vertebral body, L1-L4 lumbar vertebral body and intervertebral disc was 0.953, 0.940, 0.940 and 0.926, respectively, and the average was 0.940. Totally 197 intervertebral discs of 40 cases in test set were reconstructed with AI according to lumbar vertebrae spiral CT data. The median scores of 2 radiologists for evaluation on 197 AI reconstructed and manually reconstructed images were both 4 points, respectively, and no significant difference of the scores was found (both P>0.05). The consistency coefficient was 0.862 (95%CI 0.778, 0.946], P<0.001). Conclusion The feasibility of automatic recognition and reconstruction of intervertebral disc based on lumbar spiral CT images using convolutional neural network was very satisfied.