Convolutional Neural Networks for the Distinction of Subdural and Epidural Hematoma

Abstract

Background:Subdural and epidural hematoma are common after brain injury by the external mechanical force. The classification between subdural and epidural hematoma is mostly reckoned on the neurosurgeons’ experience in medical images. Our work proposed a model based on Convolutional Neural Networks (CNN) to improve the diagnosis accuracy of subdural hematoma and epidural hematoma. Methods: According to the inclusion criteria, totally 200 patients with subdural hematoma and epidural hematoma from the West China Hospital were enrolled. The computerized tomography (CT) images of these patients were preprocessed by mirroring, noise enhancement, increment, graying and convolution, generating 9000 original CT images. These images were divided into raining set 8100, verification set 450 and test set 450, creating the Visual Geometry Group16(VGG16) Convolutional Neural Networks. The basic model can be obtained through training Neural Network Model with data from training set. The optimal model can be obtained by modifying parameters of the basic model and optimizing it with data from validation set. With data from test set, the accuracy of the optimal model can be evaluated. Lastly, with methods such as Accuracy, Precision, Recall, F1-measure, the accuracy of this model can be evaluated. Results: The classification accuracy of the validation set on the model is 0.984 and the accuracy of CT images from the third-party organizations is 0.900. It takes 5.6 seconds averagely for the model to evaluate the image.Conclusion: The classification model has a high accuracy and short prediction time, indicating the feasibility of applying Artificial Intelligence Aided Diagnosis in practice.