U-Net模型在CT图像实现肾实质和肾窦分割及体积和径线测量

Automatic segmentation and measurement of renal parenchyma and sinus with U-Net model on CT images

目的:利用U-Net模型实现CT图像肾脏分割,测量肾实质和肾窦体积和径线。方法:搜集本院PACS中365例腹部CT增强检查中双肾正常者动脉期薄层图像。其中93例用于训练U-Net分割模型,272例用于模型效能评价。由两位影像专家检查模型返回分割结果,评价结果是否可用于体积和径线测量。以去除最小连通域方法处理图像保留像素数计算双侧肾实质和肾窦的体积。以最小体积包围盒算法测量双侧肾实质、肾窦径线。测量结果自动填写到结构化报告完成肾脏大小定量评估。建模时人工标注93例和预测时模型分割效果好272例共同用于体积和径线测量。计算肾实质、肾窦体积及三维径线95%参考值范围,采用相关性分析探讨相关因素,应用多元线性回归分析探讨其影响因素。结果:专家评价U-Net模型可很好地完成双侧肾实质和肾窦分割。测试集中分割右肾实质DICE值0.97±0.01,分割左肾实质DICE值0.97±0.01,分割右肾窦DICE值0.84±0.06,分割左肾窦DICE值0.88±0.04。多元线性回归分析显示肾实质体积=0.654×身高-0.597×年龄+0.653×体重-6.321×侧别-8.824×性别,回归方程R2为0.304;肾窦体积=0.213×体重+0.168×年龄-4.162×侧别-2.052×性别+0.122×身高,回归方程R2为0.389。模型测量结果可自动填写入结构化报告中。结论:基于U-Net可有效分割CT图像肾实质和肾窦并测量径线及体积,自动完成影像报告中双肾大小定量评估;肾实质体积和肾窦体积均与性别、年龄、身高、体重、侧别有一定关联。

Objective:To train a U-Net model for automatic segmentation of renal parenchyma and renal sinus and measurementthe volume and diameter of them.Methods:A total of 365 consecutive abdomen CT images with normal kidneys were collected retrospectively.U-Net model was trained to segment bilateral renal parenchyma and sinus.93 cases were used for modeling,and 272 cases were used for external evaluation.The segmentation results were evaluated subjectively by 2 radiologists,and Dice coefficientwas used to evaluate the results objectively.After the minimum connected domain was removed,the volume of renal parenchyma and sinus was calculated according to the number of remained pixels.The transverse(RL)diameter,anteroposterior(AP)diameter and longitudinal(SI)diameter of bilateral renal parenchyma and sinus were measured by using the algorithm of the minimum volume bounding box.The results were automatically inputted to the structured report for clinical use.The 95%reference range of bilateral renal parenchyma volumes,renal sinus volumes,and three-dimensional medians was obtained from the 365 cases.Correlation analyses were used on renal parenchyma volumes and renal sinus volumes to explore the related factors.Multivariate linear regression was performed to explore the influencing factor of them.Results:The mean Dice coefficient of held-out test dataset were 0.97±0.01,0.97±0.01 for the right and left renal parenchymal segmentation,and 0.84±0.06,0.88±0.04 for the right and left renal sinus segmentation.Multiple linear regression analysis showed that renal parenchymal volume=0.654×height-0.597×age+0.653×weight-6.321×lateral-8.824×sex(R2=0.304).Renal sinus volume=0.213×weight+0.168×age-4.162×side-2.052×sex+0.122×height(R2=0.389).Conclusion:The U-Net is good for renal parenchyma and renal sinus segmentation and can give the measurement automatically to a structured report.The volume of the renal parenchymal and sinus was related to gender,age,height,weight,and side in normal individuals.