基于CT平扫的深度学习自动分割模型对肾积水病人分侧肾功能的评估

Assessment of split renal function in hydronephrosis patients with automatic segmentation model based on CT plain scan and deep learning

目的探讨基于CT平扫的深度学习自动分割模型对肾积水病人分侧肾功能评估的价值。方法回顾性收集2所医院共209例肾积水病人的平扫CT影像、年龄、性别、体质量指数(BMI)以及基于单光子发射计算机体层成像(SPECT)测量的肾脏肾小球滤过率(GFR),并以其来源医院确定为训练集(137例)和测试集(72例)。采用U-Net方法构建肾脏自动分割模型,用于自动分割平扫CT影像上肾积水和肾实质区域,计算肾积水、肾实质体积及两者的体积比作为肾脏形态特征。根据GFR值将肾功能分为正常[GFR≥30 mL/(min·1.73 m^(2))]与异常[GFR<30 mL/(min·1.73 m^(2))]。使用多因素逻辑回归筛选独立预测特征并建立分侧肾功能评估模型。采用Dice相似性系数(DSC)评价自动分割结果,采用受试者操作特征曲线下面积(AUC)评价分侧肾功能评估模型的效能,使用DeLong检验比较AUC的差异。结果影像采用自动分割平均耗时为每例病人2.2 s,而手动分割耗时是自动分割的671.8倍。肾实质和肾积水自动分割的平均DSC分别为0.89和0.63。基于肾积水和肾实质的体积比、肾实质体积、年龄、BMI构建的分侧肾功能评估模型在测试集上的AUC为0.809。基于自动分割与手动分割的分侧肾功能评估模型效能差异无统计学意义(P>0.05)。结论基于CT平扫和深度学习的肾脏自动分割模型在肾积水病人分侧肾功能评估中具有较好的价值,有望提高诊断效率。

Objective To explore the value of an automatic kidney segmentation model based on plain CT scan and deep learning in the assessment of split renal function in patients with hydronephrosis.Methods The plain CT images,age,sex,body mass index(BMI)and glomerular filtration rate(GFR)of each kidney measured by SPECT in 209 patients with hydronephrosis in 2 hospitals were collected retrospectively,and determined as the training set(137 cases)and the test set(72 cases)by the source hospital.The U-Net method was used to construct an automatic kidney segmentation model for segmentation of hydronephrosis and renal parenchyma on plain CT images.The volumes of hydronephrosis and renal parenchyma and the ratio of the two volumes were calculated,and treated as the kidney morphological features.The kidneys were divided into two types:normal[GFR≥30 mL/(min·1.73 m^(2))]and impaired[GFR<30 mL/(min·1.73 m^(2))].Multivariable logistic regression was used to select independent predictive features and construct a split renal function assessment model.The Dice similarity coefficient(DSC)was used to evaluate the results of automatic segmentation,the area under the receiver operating characteristic curve(AUC)was used to evaluate the performance of the split renal function assessment model,and the DeLong test was used to compare the differences in AUC.Results The average time spent on kidney automatic segmentation was 2.2 seconds/case.The manual segmentation took 671.8 times longer than automatic segmentation.The mean DSC of the automatic segmentation of renal parenchyma and hydronephrosis were 0.89 and 0.63.Based on the volume ratio of hydronephrosis and renal parenchyma,renal parenchymal volume,age,and BMI,the AUC of the split renal function assessment model on the test set was 0.809.There was no significant difference in the performance of the split kidney function assessment based on automatic and manual segmentation(P>0.05).Conclusion The kidney automatic segmentation model based on CT plain scan and deep learning has good performance in the assessment of split renal function status in patients with hydronephrosis,which is expected to improve the efficiency of clinical diagnosis.