Research On Multi-Modal Brain Image Recognition Method Based On Deep Learning

Alzheimer’s disease(AD)is a degenerative disease of the central nervous system.Once the disease is irreversible,it can only be procrastinated by drugs,causing distress and threats to the patient’s life and even life.Mild Cognitive Impairment(MCI)is a clinical state between normal aging and dementia,and its risk of converting into AD is about 10 times that of normal elderly people.MRI and PET are commonly used neuroimaging methods in AD and MCI research.This article uses deep learning to extract hidden features from neuroimaging for classification and identification of AD and MCI patients.The specific work of this article is as follows:Select the neuroimaging data in the ADNI database as the experimental research data in this paper,screen the subjects with both MRI and PET modalities,and perform preprocessing operations on the three-dimensional original MRI brain image,including segmentation,standardization and smoothing to obtain gray matter images.The original three-dimensional PET images are smoothed and standardized,and then slicing the two modal pre-processed images.The processed data can be used as input data for the deep learning network model.An improved MRI brain image classification method based on improved topological sparse coding is proposed.This method uses the improved topological sparse coding network model to construct a deep neural network,and uses a greedy algorithm to iteratively optimize the cost function.The weight matrix learned by the network model is the visual feature of the brain image.An improved L-BFGS sparse denoising auto-encoder network is proposed and applied to MRI brain image recognition.This method combines two algorithms(greedy algorithm and L-BFGS)to iteratively optimize the sparse denoising auto-encoder network model,extracts features of MRI brain gray matter images and classify.An adaptive weighted integrated convolutional neural network is proposed and applied to MRI brain image recognition.This method trains the convolutional neural network by using brain images with different cross sections of the same brain,extracts the classification prediction feature values of the softmax layer,and then the adaptive weighting method is used to perform weighted voting on the three cross-section data(coronal,sagittal and axial),and finally the integrated learning technology is used toperform weighted integration on the data of different cross sections to obtain the classification accuracy rate.A heterogeneous multimodal brain image recognition method based on a convolutional neural network is proposed.This method uses a convolutional neural network to extract pre-processed fully connected layer brain image features,and then two different modalities of the same individual are spliced and integrated,finally the supervised learning method and the unsupervised learning method are used to classify the integrated features to obtain the final recognition accuracy rate.