基于深度学习的自发性脑出血CT影像分割算法精准计算病灶体积的应用探讨  

Evaluation of spontaneous intracerebral hemorrhage by using CT image segmentation and volume assessment based on deep learning

作　　者：王霁雯 林雨[1] 熊建华[1] 于圣平[1] 魏伟[1] 杨新宇[1] 肖福顺[1] 王永利[1] 梁孔明 王浩 李秀丽 刘兵 Wang Jiwen;Lin Yu;Xiong Jianhua;Yu Shengping;Wei Wei;Yang Xinyu;Xiao Fushun;Wang Yongli;Liang Kongming;Wang Hao;Li Xiuli;Liu Bing(Department of Neurosurgery,General Hospital,Tianjin Medical University,Tianjin 300052,China;Deepwise Artificial Intelligence Lab,Beijing 100080,China)

机构地区：[1]天津医科大学总医院神经外科,300052 [2]深睿医疗人工智能研究院,北京100080

出　　处：《中华放射学杂志》2019年第11期941-945,共5页Chinese Journal of Radiology

摘　　要：探讨深度学习技术在自发性脑出血CT影像分割和精准计算出血病灶体积的效果.方法回顾性分析天津医科大学总医院2016年4月至2018年4月影像及临床诊断为自发性脑出血的1223例患者资料.分为实质内出血、脑室出血、蛛网膜下腔出血和混合型出血4种类型.采用完全随机法将患者分为训练集905例、验证集156例、测试集162例,实质内出血分别为498、107和100例.由医师对出血区域进行轮廓勾勒标注,作为金标准构建模型以评测测试集性能.采用脑卒中人工智能检测分析系统构建模型,并采用多田公式计算出血体积.在测试集中筛选97例单纯实质内出血患者,按照实质内出血体积,将97例患者分为<5 ml组、5~25 ml组及>25 ml组,记录采用多田公式、模型预测的出血体积绝对和相对误差,并记录模型预测的Dice指数.结果测试集162例中,深度分割模型在实质内出血、脑室出血、蛛网膜下腔出血和混合型出血的Dice指数分别为0.87、0.85、0.67和0.77.在单纯实质内出血的97例患者中,模型计算血肿量为(29.55±37.69)ml,多田公式计算的血肿量为(24.04±31.22)ml.实质内出血体积<5 ml组、5~25 ml组及>25 ml组,采用模型预测的出血体积绝对误差分别为(0.52±0.54)、(1.53±1.22)、(7.93±8.49)ml,多田公式得到的出血体积绝对误差分别为(0.68±0.60)、(3.16±2.90)、(19.31±17.23)ml.结论利用深度学习模型自动分割脑出血可以应用于实质内血肿计算,误差小于多田公式计算结果.Objective To evaluate the feasibility and accuracy of deep learning in CT image segmentation and further lesion-volume assessment of spontaneous intracerebral hemorrhage.Methods A total of 1223 cases of spontaneous intracerebral hemorrhage including parenchymal hemorrhage,ventricular hemorrhage,subarachnoid hemorrhage and mixture hemorrhage,from April 2016 to April 2018 in Tianjin Medical University General Hospital,were retrospectively enrolled and analyzed.The patients were randomly divided into training set(905 cases),validation set(156 cases)and test set(162 cases),among each group,the number of parenchymal hemorrhage was 498,107 and 100,respectively.The bleeding area manually outlined by physician was served as the reference standard to build the segmentation model and to evaluate the performance of the validation set.Patients were divided into 3 groups according to the volume calculated by reference standard.The volume of hematoma in group 1 was less than 5 ml,while group 2 was 5-25 ml,and group 3 was more than 25 ml.Comparison of the hematoma volume calculated by segmentation model and that calculated by ABC/2 formula was conducted in 97 simple intraparenchymal hemorrhage cases.Results In 162 cases of test set,the Dice coefficients of the segmentation model were 0.87,0.85,0.67 and 0.77 in parenchymal hemorrhage,intraventricular hemorrhage,subarachnoid hemorrhage and mixture hemorrhage,respectively.The estimated hematoma volume in the 97 intraparenchymal hemorrhage cases calculated by the segmentation model was(29.55±37.69)ml,and that calculated by the ABC/2 formula was(24.04±31.22)ml.Compared with reference standard,the absolute errors of three segmentation model were(0.52±0.54),(1.53±1.22)and(7.93±8.49)ml in group 1,2 and 3 respectively.The absolute errors of the ABC/2 formula were(0.68±0.60),(3.16±2.90)and(19.31±17.23)ml in group 1,2 and 3.Conclusion Deep learning based segmentation model improved detection of intraparenchymal hematoma volume,compared with ABC/2 formula.

关 键 词：深度学习 图像分割 脑出血 

分 类 号：R74[医药卫生—神经病学与精神病学]