CT Imaging Analysis Of Intracranial Hemorrhage Based On Deep Learning

Intracranial hemorrhage is a high incidence rate and high mortality rate of cerebrovascular disease.Accurate clinical diagnosis and timely intervention plan will help to improve the survival rate of patients.CT is the first choice for the initial diagnosis of intracranial hemorrhage.The study of CT image analysis of intracranial hemorrhage based on deep learning will help to reduce the missed diagnosis rate and misdiagnosis rate of doctors.The analysis of intracranial hemorrhage mainly includes two tasks:classification of intracranial hemorrhage subtype and segmentation of intracranial hemorrhage region.In the task of subtype classification of intracranial hemorrhage,most researches are based on convolutional neural network,but the single prediction probability output by these methods is difficult to determine the reliability of the model.How to obtain a safe and reliable deep learning model is a very challenging task.In the task of intracerebral hemorrhage region segmentation,traditional methods propose different manual features for different types of intracerebral hemorrhage.However,these carefully designed manual features are difficult to adapt to different types of intracerebral hemorrhage segmentation at the same time.Deep learning can solve this problem through strong representation ability.However,the segmentation of intracerebral hemorrhage region based on deep learning still faces the following problems Problem: it takes a lot of time and energy for human experts to segment and label intracranial hemorrhage,so it is difficult to obtain a large number of labeled data;deep learning model often has a large number of parameters,the lack of a large number of labeled data will make the model suffer from the risk of over fitting,so it limits the performance of deep learning in the segmentation of intracranial hemorrhage.In view of the problems existing in the analysis of intracranial hemorrhage in the above-mentioned deep learning,this paper proposes the following methods to deal with these challenges:· Based on the current advanced convolutional neural network,combined with the method of estimating model uncertainty in the Bayesian deep learning framework,it is used in the task of classification of intracranial hemorrhage subtypes,and at the same time,the influence area of model prediction is visualized by category activation map.Model uncertainty estimation and visualization can be used to help determine whether the prediction of the model is reliable.In addition,it can also be used to screen model prediction and guide human to find neglected bleeding areas.Sufficient experiments show that this method has excellent performance in the classification of intracranial hemorrhage and its subtypes.In addition,uncertainty estimation can be effectively used to judge the reliability of model prediction.· Based on the teacher student model,the observation uncertainty estimation is introduced to the semi supervised segmentation learning of intracranial hemorrhage.The observation uncertainty estimation can capture the ambiguous regions in CT images,and select the regions with high reliability in the consistency regularization,so that the model can use the dimensionless data more effectively.Through sufficient experiments,it is proved that this method can effectively improve the performance of intracranial hemorrhage region segmentation by using non labeled CT images.· In this paper,the capsule network is applied to the segmentation of intracranial hemorrhage area,and the grouping capsule structure is proposed.This structure can effectively solve the problems of high memory occupation and computational complexity caused by capsule computing.Compared with the traditional convolution neural network,capsule computing can effectively keep the equivariant feature.The experimental results show that the network can significantly reduce the time of capsule calculation and achieve good performance in the segmentation of intracranial hemorrhage area.