Research On Subtype Classification And Segmentation Algorithm Of Intracranial Hemorrhage Based On CT Images

Intracranial hemorrhage is a kind of cerebrovascular disease with high morbidity.Accurate diagnosis and timely treatment are helpful to improve the recovery rate of patients.The computer-aided diagnosis and treatment system for intracranial hemorrhage is very important for doctors’ diagnosis and follow-up treatment.At present,the research of computer-aided diagnosis and treatment can be divided into two fields: classification of intracranial hemorrhage subtypes and segmentation of intracranial hemorrhage lesions.In the classification of intracranial hemorrhage subtypes,most studies use the network architecture combined with CNN-RNN.However,these studies have ignored the following problems :(1)the existing CNN networks,such as RESNET and DENSENET,are not capable of feature expression in CT images;(2)The proportion of positive and negative samples in the data set of intracranial CT images is seriously unbalanced;(3)There may be one or more types of lesions in a single hemorrhage CT section,and the correlation between these lesions is often neglected.On the other hand,in the field of intracranial hemorrhage segmentation,the U-NET segmentation method has shown significant performance advantages.However,the network design of U-NET still has the following problems :(1)the convolution operation of the coding layer of U-NET is too simple,and it cannot extract more refined features;(2)The jump connection design of U-NET enables the feature map of the shallow coding layer to be directly connected with the feature map of the deep decoding layer,which means that the decoding layer will introduce irrelevant noise and lead to the misclassification of pixels;(3)The deconvolution operation of the decoding layer cannot gather the multi-scale feature information,thus affecting the restoration of the edges and details of the image focus,and limiting the performance of the network segmentation.For the unsolved problems in the above two areas,this paper proposes the following methods to deal with these challenges:1.The problem of intracranial hemorrhage subtype classification was transformed into a fine-grained classification problem,and a compact linear pooling method was introduced.Through bilinear convergence of features,the CNN pathway could extract more refined image features;To solve the problem of unbalanced data set distribution,multi-weighted Focal(MWF)loss was proposed.This paper uses dynamic weight coefficients to control the prediction loss of positive and negative samples to balance the parameter shift in the network learning process.A Softmax Relative Entropy(SRE)loss was proposed to measure the correlation between multiple label categories by calculating the difference between the true label distribution and the predicted label probability distribution.2.A new network framework,Multiscale U-NET(MU-NET),is proposed and applied to the segmentation of intracranial hemorrhage lesions.In the encoder module,the network module of Res2 Net is introduced.This design can extract more fine multi-scale features and increase the receptive field of the feature map.In order to reduce the semantic gap between the corresponding layers of the coding layer and the decoding layer,a Multi-Encoder Information Fusion(MIF)module is proposed to make up for the global Information lost by the decoder through Information Fusion.In addition,in order to gather feature information at multiple scales to better restore the edges and details of lesions,a new decoder module called Multiscale Deconvolution Block(MDB)is proposed.A variety of experiments show that the proposed method is effective and surpasses many advanced methods in performance.