颅脑CT影像深度学习预测脑出血破入脑室

Prediction of intraventricular hemorrhage based on deep learning of brain CT images

目的 探索深度学习技术在脑出血是否破入脑室自动分类方面的应用。方法 收集2010年1月到2020年12月陆军军医大学第一附属医院神经外科收治的1 027例自发性脑出血患者的颅脑CT影像,将每层图像划分到正常、未破入脑室、破入脑室3个类别,利用DenseNet121、ResNet50、ResNet101、Swin-base、Vit-base与VGG16等6种典型的深度网络分别构建用于判断脑出血是否破入脑室的分类模型,并在内部数据集和外部数据集(CQ500)上分别进行测试。为增强深度学习网络的可解释性,利用EigenGradCAM方法制作热力图对深度模型的关注区域进行可视化。结果 利用精确率、召回率、特异性、阴性预测值与F1值评价深度模型性能,VGG16模型在内部测试集上,正常组分别取得0.983、0.977、0.984、0.978与0.980,未破入脑室组分别取得0.917、0.902、0.965、0.958与0.909,破入脑室组分别取得0.877、0.911、0.966、0.976与0.894;外部测试集上,正常组分别取得0.967、0.870、0.985、0.938与0.916,未破入脑室组分别取得0.827、0.939、0.902、0.967与0.879,破入脑室组分别取得0.938、0.906、0.970、0.954与0.922;内部测试集和外部测试集的准确率分别为0.940、0.905。基于EigenGradCAM方法制作的热力图表明VGG16能够合理关注到相关区域。结论 利用VGG16构建的深度模型在判断脑出血是否破入脑室方面取得了最优的预测性能,表明深度学习可以应用于脑室出血的判断。

Objective To explore the application of deep-learning technology in automatic classification of intraventricular hemorrhage.Methods The brain CT images of 1 027 patients with spontaneous ICH from the First Affiliated Hospital of Army Medical University from January 2010 to December 2020 were collected and retrospectively analyzed,which were subsequently divided into 3 types:normal,intracerebral hemorrhage with/without intraventricular extension.Six typical deep networks,including DenseNet121,ResNet101,ResNet50,Swin-base,Vit-base and VGG16,were used to construct classification models for identifying whether there was intraventricular extension of intracerebral hemorrhage.The performance of each model was assessed on both the internal and external data sets(CQ500) respectively.In order to enhance the interpretability of the deep learning networks,the EigenGradCAM method was adopted to generate heatmaps for visualization of the interested regions of the model.Results The performance of the models was evaluated based on precision,recall rate,specificity,negative predictive value(NPV) and F1 value.On the internal test set of VGG16 model,the normal group achieved 0.983,0.977,0.984,0.978 and 0.980,respectively;the intracerebral hemorrhage without intraventricular extension group achieved 0.917,0.902,0.965,0.958 and 0.909;and the intraventricular extension of intracerebral hemorrhage group obtained 0.877,0.911,0.966,0.976 and 0.894,respectively.Meanwhile,on the external test set of VGG16 model,the normal group reached 0.967,0.870,0.985,0.938 and 0.916;the intracerebral hemorrhage without intraventricular extension group reached 0.827,0.939,0.902,0.967 and 0.879;and the intraventricular extension of intracerebral hemorrhage group achieved 0.938,0.906,0.970,0.954 and 0.922.The accuracy(ACC) of internal and external test sets was 0.940 and 0.905,respectively.Finally,the heatmaps generated by EigenGradCAM method showed the VGG16 could reasonably focus on the relevant areas.Conclusion The VGG16 achieved the best predictive performance for identifying whether there is an intraventricular extension in intracerebral hemorrhage or not,indicating that deep learning can be effectively applied to the judgement of intraventricular hemorrhage.