摘要
目的:开发一种可以检测不同类型颅内出血并自动计算血肿体积的基于卷积神经网络的深度学习算法,探讨其识别的准确性及血肿分割的一致性。方法:数据集1纳入9594例颅脑CT平扫图像,随机选取223例颅内出血阳性患者作为颅内出血类型识别的测试集,剩余CT图像作为其训练集,评估测试集中算法识别五种不同类型颅内出血的效能。数据集2选取另外819例已人工勾画出血灶的CT图像,随机选取74例作为测试集,以人工手动分割为金标准,验证测试集中算法分割与人工分割的一致性。结果:在223例颅内出血阳性患者中,深度学习算法对五种类型颅内出血识别的曲线下面积均大于或接近0.85,特异度均大于0.95;在74例血肿分割测试数据中,算法自动测量的血肿体积与人工手动分割测量的血肿体积之间达到较高的一致性,脑实质内出血、硬膜外出血、脑室内出血及硬膜下出血体积测量的组内相关系数分别为1、0.990、0.996和0.878。结论:基于卷积神经网络的深度学习算法可以较好地识别不同类型的颅内出血,并能精确测量血肿体积,具有一定的临床应用前景。
Objective:To develop a deep learning algorithm based on convolutional neural network to detect different types of intracranial hemorrhage and automatically calculate the volume of hematoma.And to explore its accuracy and consistency of hematoma segmentation.Methods:9594 cases of non-contrast head CT scan images were retrospectively included in Dataset 1.223 positive cases were randomly selected as the testing dataset for intracranial hemorrhage subtype recognition,and the rest CT images were used as the training dataset.The effectiveness of the test set algorithm in recognizing five different types of intracranial hemorrhage were evaluated.Another 819 manually labeled CT scans were enrolled in Dataset 2,from which 74 cases were randomly selected as the testing dataset.Manual segmentation was taken as the gold standard to verify the consistency of algorithm segmentation and manual segmentation in the test set.Results:Among 223 patients with positive intracranial hemorrhage,the area under the curve(AUC)for five types of intracranial hemorrhage in the subtype recognition testing dataset were all above or approximate to 0.85,and the specificity were greater than 0.95.Among the 74 cases of hematoma segmentation test data,there was a high consistency between the volume of hematoma automatically measured by the algorithm and the volume of hematoma manually measured by the algorithm.The intraclass correlation coefficients of intraparenchymal hemorrhage,epidural hemorrhage,intraventricular hemorrhage and subdural hemorrhage were 1,0.990,0.996 and 0.878 respectively in the hematoma segmentation testing dataset.Conclusion:Deep learning algorithm based on convolution neural network has good performance in intracranial hemorrhage recognition and hematoma segmentation,which has a certain prospect of clinical application.
作者
李娟
汤翔宇
沈逸
廖术
石峰
朱文珍
LI Juan;TANG Xiang-yu;SHEN Yi(Department of Radiology,Tongji Hospital,Tongji Medical College,Huazhong University of Science and Technology,Wuhan 430030,China)
出处
《放射学实践》
CSCD
北大核心
2021年第1期7-12,共6页
Radiologic Practice
基金
国家自然科学基金青年基金(51907077)。
关键词
卷积神经网络
深度学习
颅内出血
血肿分割
体层摄影术
X线计算机
Convolution neural network
Deep learning
Intracranial hemorrhage
Hematoma segmentation
Tomography,X-ray computer
