Research On Automatic Classification And Extended Prediction Of Cerebral Hemorrhage Based On Deep Learning

Acute intracerebral hemorrhage is the cerebrovascular disease with the highest mortality rate,and early diagnosis and expanded prediction of the type of intracerebral hemorrhage help to reduce patient mortality.With the development of computer technology and the improvement of CT imaging technology,the amount of medical imaging data has shown an explosive growth,traditional methods that rely on doctors’ manual marking and feature extraction are difficult to meet the needs of massive data analysis,and a long time of work will lead to doctors’ misdiagnosis.Therefore,there is an urgent need for computer-assisted technology to reduce work pressure for doctors and provide a basis for objective diagnosis.This paper proposes a new computer-aided diagnosis and prediction method based on deep learning technology,uses the powerful feature learning ability of deep learning,explores the distributed representation of different levels of data,and focuses on the automatic classification and expansion prediction methods of cerebral hemorrhage based on deep convolutional neural networks.The work of this article mainly includes the following two parts:(1)Aiming at the problem of low degree of fit of the segmentation edge of cerebral hemorrhage and low accuracy of classification of cerebral arteriovenous malformations,a method of cerebral hemorrhage image classification with shared shallow parameter multi-task learning is proposed.CT images play an important role in the timely diagnosis of patients with suspected acute cerebral hemorrhage.Medical staff usually use CT images of patients to manually segment and classify bleeding lesions.The process is complicated and the classification accuracy is low.In order to shorten the diagnosis time and improve the accuracy of cerebral hemorrhage classification,this paper proposes a new computer-aided diagnosis method,which is a method of cerebral hemorrhage image classification that shares shallow parameter multi-task learning.First,build a segmentation sub-module based on the 3D U-Net network and a classification sub-module based on a convolutional neural network,sharing the shallow parameters of the two sub-modules,extract the public features between two related tasks in the shallow layer of the network,and extract the private features of the two tasks in the deep layer of the network.At the same time,the model uses a composite loss function according to the difficulty of the task.By adjusting the weight of the composite loss function,the model achieves good segmentation and classification performance.The segmentation average dice similarity coefficient of this method on the validation set reaches 0.828,the Hausdorff distance is 9.895,and the classification accuracy,sensitivity and specificity reach 95.0%,90.5% and 100.0%,respectively.Experimental results prove that the method proposed in this paper can not only improve the fitting degree of cerebral hemorrhage,but also improve the accuracy of cerebral hemorrhage classification,effectively assist doctors in locating the lesion,and diagnose the type of cerebral hemorrhage in time,which is of great significance to the research of auxiliary diagnosis of cerebral hemorrhage.(2)In view of the obvious differences in the baseline volume of hypertensive intracerebral hemorrhage,which is the most difficult to treat,and the complicated process of traditional expansion risk prediction methods,a multi-scale feature fusion of the attention mechanism is proposed.The brain hemorrhage enlargement risk prediction model based on multi-scale feature fusion of attention mechanism is a computer-aided diagnosis method based on deep learning.The model consists of three parts: based on the main part of the residual network,it can maintain a strong expressive ability in the deep network,and fully extract CT image features;the binary segmentation auxiliary task based on spatial upsampling is used as the attention module to make the whole focus of the network is concentrated on the bleeding lesion area;the multi-scale feature fusion module based on multi-core pooling can make full use of the picture spatial context information.On the hypertensive intracerebral hemorrhage data set,the classification accuracy of this method reached 88.34%.The experimental results show that the multi-scale feature fusion model of the attention mechanism can effectively improve the predictive performance of the risk of hypertensive cerebral hemorrhage.To sum up,this article first uses the shared shallow parameter multi-task learning method to classify cerebral hemorrhage CT images,quickly and accurately identify cerebral arteriovenous malformation hemorrhage and other types of hemorrhage.Then,for the most difficult to treat hypertensive cerebral hemorrhage among other types of hemorrhage,a multiscale feature fusion network of attention mechanism was constructed to expand risk prediction and provide objective advice to doctors.