Research And Implementation Of HAND Early ANI Diagnosis Based On Multimodal Fusion

AIDS(acquired immune deficiency syndrome,abbreviated as AIDS)is a malignant infectious disease,which is caused by human immunodeficiency virus(HIV)invading the human body,damaging the human immune system,and finally leading to death,which greatly threatens the health of people all over the world.For the high-risk group of HIV infection,we should actively detect;for the confirmed infected group,we should early diagnosis and treatment,and effectively block the progress of the disease.In the early stage of HIV infection(1 week),macrophages infected with HIV enter the nerve center through the blood-brain barrier,causing chronic long-term damage to nerve cells.Even after antiretroviral treatment,at least 50%of HIV infected people still suffer from sensory and motor dysfunction and neurocognitive impairment,which is HIV associated neurocognitive disorders,HAND).The evolution of HAND is from asymptomatic neurocognitive impairment(ANI),mild neurocognitive disorder(MND),to HIV associated dementia,There are three stages of had,so if we can identify the stage of ani effectively,it is of great significance for the early diagnosis and treatment of HAND.In this paper,a multimodal model based on convolutional neural network is proposed to assist the diagnosis of ANI.This paper selects MRI,DTI and rs-fMRI images of ANI patients from Beijing You’an Hospital Affiliated to Capital Medical University and normal control group,and builds an efficient and accurate auxiliary diagnosis system through neural network based on multimodal thinking.In order to solve the problem that the quality of hand samples is different greatly and the quantity is less,this paper puts forward the data pre-processing measures,including format,calibration,data availability enhancement and sample balance.Through these measures,the diversity of samples can be reconstructed to achieve data balance.Aiming at the problems of low classification accuracy and poor operation efficiency of single-mode data using 3D CNN network,this paper builds a 3D CNN as the main body of the Incaption-ResNet structure classification model,this model can make full use of its 3D characteristics to learn the spatial information contained in the MRI image,the concept RESNET structure can effectively extract the features of different levels of samples,at the same time,prevent the gradient disappearance caused by the deeper network layers and the problem of model over fitting,and achieve the accuracy of the same-layer 3D CNN model increased by 9%and training time reduced by 212 minutes.To solve the problem that the fusion scheme is not clear in multi input scenarios,this paper compares six fusion classification schemes,proposes a multi-channel feature fusion algorithm based on attention mechanism,selects the brain features extracted based on all template for Fisher feature selection,and finally uses attention mechanism with the features extracted from end-to-end to give different weight to different feature paths,and finally the number in ani On the data set,the accuracy of multi-channel attention mechanism based multi-feature fusion scheme is 6.5%higher than that of single mode,5.9%higher than that of end-to-end feature fusion scheme,and 3.2%higher than that of average weight fusion.