基于改进U-net网络的脑血肿CT图像分割与体积测量研究

Segmentation and volume measurement of intracerebral hemorrhage in CT images based on improved U-NET network

目的对U-net网络进行改进,构建相对泛化的深度学习算法,分割及精准评估脑血肿体积以减少人工计算误差,为临床提供新的计算机辅助诊断工具。方法回顾性收集来自2家医院的242例幕上自发性脑出血CT图像以及临床资料,随机分为训练集(含验证集)192例、测试集50例。基于改进的U-net网络,构建了MSHA-Unet模型对脑血肿进行分割,以人工手动分割血肿体积为标准,对比新模型与多田公式在血肿体积测量中的准确性、稳定性。结果MSHA-Unet深度学习算法血肿分割模型分割血肿体积值更接近人工手动分割结果,与多田公式测量结果相比,新模型算法分割血肿体积测量的百分误差较小,为8.197%,组内相关系数为0.987,96%(48/50)的数据在95%一致性界限(LoA),并且其95%LoA较窄,为1.70~4.06 mL。结论MSHA-Unet网络模型能够较好地解决脑血肿的分割问题,利用数据集建立分割模型,是减少人力成本构建更具泛化能力模型的可行办法。

Objective To improve the U-net network and construct a relatively generalized deep learning algorithm to segment intracerebral hematoma and accurately evaluate the volume to reduce the manual calculation error,and to provide a new computer-aided diagnostic tool for the clinic.Methods Retrospective collection of CT images and clinical data of 242 episodic spontaneous cerebral hemorrhage cases from 2 hospitals were randomly divided into a training set(including validation set)of 192 cases and a test set of 50 cases.Based on the improved U-net network,the MSHA-Unet model was constructed to segment the intracerebral hematoma,and the accuracy and stability of the new model and the Tada formula in hematoma volume measurement were compared with manual manual segmentation of the hematoma volume as the standard.Results MSHA-Unet deep learning algorithm segmentation of hematoma volume was closer to the manual segmentation results.Compared with the measurement results of Tada formula,the error of the new model is smaller,which was 8.197%.An intra-group correlation coefficient of 0.987,96%(48/50)of the data at the 95%limits of agreement(LoA),and its 95%LoA was narrower,ranging from 1.70 to 4.06 mL.Conclusion The MSHA-Unet network model can better solve the segmentation problem of intracerebral hematoma,and using the dataset to build the segmentation model is a feasible way to reduce the labor cost to build a more generalizable model.