利用深度学习实现CT图像上腰骶椎各结构分割及椎间盘自动定位的可行性研究

Feasibility study on the segmentation of lumbosacral vertebral structures and the automatic localization of intervertebral disc in CT images by deep learning

目的:探索使用3D U-Net深度学习模型在CT图像上进行腰椎各结构的自动分割及椎间盘自动定位的可行性。方法:回顾性搜集2020年12月1日-2021年3月29日于本院行腰椎平扫CT的患者影像图像,排除腰椎术后、脊柱畸形及骨转移的病例,共纳入了154个图像数据。手工标注腰骶椎各椎体、椎间盘及硬膜囊。按8:1:1比例将数据随机分为训练集(n=125)、调优集(n=14)和测试集(n=15)。利用3D U-Net分割模型进行训练,以医师手动标注结果作为参考标准,根据测试集Dice相似系数(DSC)、体积相似度(VS)和Hausdorff距离(HD)作为评价模型分割效能的指标。应用连通域分割算法进行腰椎各椎间盘定位,以医师判定为金标准,采用混淆矩阵评价模型识别各椎间盘的位置的定位效能。结果:测试集中3D U-Net深度学习模型对腰骶椎各结构分割结果DSC值、VS值均>0.96。模型识别各椎间盘位置的准确率达98.7%,模型预测与医师判定一致性高。结论:3D U-Net深度学习模型和可用于CT图像中腰椎各主要结构的自动分割并通过连通域算法实现椎间盘自动定位。

Objective:To explore the feasibility of using U-Net deep learning model for automatic segmentation of various lumbar structures and the automatic localization of intervertebral discs on CT images.Methods:From December 1,2020,to March 29,2021,the images of patients who underwent CT scans of lumbar spine in our hospital were retrospectively collected,excluding the cases with lumbar surgery,spinal deformity and bone metastasis,and a total of 154 data were included.The lumbosacral vertebral bodies,intervertebral discs and dural sacs were manually marked.According to the ratio of 8:1:1,the data were randomly divided into training set(n=125),validation set(n=14)and test set(n=15).The segmentation model was trained by 3U-Net model,and the results of manual labeling by physicians were used as the reference standard.The dice similarity coefficient(DSC),volume similarity(VS)and hausdorff distance(HD)were used as the indexes to evaluate the segmentation efficiency of the model.With physician's judgment as the gold standard,the confusion matrix was used to evaluate the positioning efficiency of the model in identifying the location of each disc.Results:In the test set,the DSC and VS values of the segmentation results of lumbosacral vertebral structures by 3D U-Net deep learning model were all greater than 0.96.The accuracy of the model in identifying the location of each disc was 98.7%,and the model prediction was in good agreement with the physician's judgment.Conclusion:The 3D U-Net deep learning model can be used to automatically segment the main structures of lumbar vertebrae in CT images and realize automatic disc localization through connected domain algorithm.