基于V-Net卷积神经网络深度学习模型自动分割腰椎CT图像中的椎旁肌

目的 观察基于V-Net卷积神经网络(CNN)的深度学习(DL)模型自动分割腰椎CT图像中的椎旁肌的价值。方法 收集471例接受腰椎CT检查患者，按7∶3比例将其分为训练集(n=330)和测试集(n=141);采用2D V-Net进行训练，建立DL模型；观察其分割腰大肌、腰方肌、椎后肌群及椎旁肌的价值。结果 基于V-Net CNN的DL模型分割椎旁肌精度良好，戴斯相似系数(DSC)均较高、肌肉横截面积误差率(CSA error)均较低；其分割训练集图像中的腰大肌、腰方肌及椎旁肌的DSC均高于测试集(P均<0.05),而分割训练集中4组肌肉的CSA error均低于测试集(P均<0.05)。测试集内两两比较结果显示，该模型分割椎后肌群的DSC最高、腰方肌的DSC最低；分割腰方肌的CSA error最高、椎旁肌的CSA error最低(P均<0.05)。结论 以基于V-Net的DL模型自动分割椎旁肌的效能较佳。

Deep learning algorithm based on V-Net convolutional neural network for automatic segmentation of paraspinal muscles on lumbar CT images

Objective To observe the value of deep learning (DL) algorithm based on V-Net convolutional neural network (CNN) for automatic segmentation of paraspinal muscles on lumbar CT images. Methods Totally 471 patients who underwent lumbar CT examination were enrolled and divided into training set (n=330) or test set (n=141) at the ratio of 7:3. 2D V-Net was used for training, then DL algorithm was established, and its value for segmentation of psoas major, quadratus lumborum, posterior muscle groups and paraspinal muscles were observed. Results DL algorithm based on V-Net CNN had good accuracy for segmentation of paraspinal muscles, with high Dice similarity coefficient (DSC) and low muscle cross-sectional area error (CSA error). DSC of segmentation of psoas major, quadratus lumborum and paraspinal muscles in training set were higher than those in test set (all P<0.05), while CSA error of segmentation of 4 groups of muscles in training set were all lower than those in test set (all P<0.05). Pairwise comparison was performed in test set, and the results showed that DSC of posterior muscle groups was the highest, and that of quadratus lumborum was the lowest, while CSA error of quadratus lumborum was the highest and that of paraspinal muscles was the lowest (all P<0.05). Conclusion DL algorithm based on V-Net CNN had good value for automatic segmentation of paraspinal muscles.