A Multimodal Neuroimaging-based Study To Predict The Progression Of Mild Cognitive Impairment

Alzheimer’s disease(AD)is a classic neurodegenerative disease with an unknown cause and an irreversible course to date.Mild cognitive impairment(MCI)is a transitional state between AD and healthy individuals.MCI can be divided into two categories: progressive MCI(p MCI),which will develop into AD in a short period of time,and stable MCI(p MCI),which will not develop into AD within 4 years.It is important to predict the development of MCI during the period of MCI and to intervene to prevent the development of AD in high-risk types.Currently for prediction of MCI patients,the clinicians rely on physicians to make predictions from the patients’ periodic reexamination results.The input data types of existing prediction methods on MCI patients are: unimodal,multimodal,and multiparametric,and most of the methods are based on traditional machine learning,including longitudinal analysis based on time-progressed data sets.Multimodality contains more feature information compared to unimodality,and multiparameter refers to clinical parameters related to the onset of AD.Although the above methods have achieved certain prediction results,there are problems such as the number of samples meeting the requirements is less compared to natural images and the model is less robust.Based on the above problems,this study used 3D-DenseNet for the prediction work of MCI patients,and used it for feature extraction of magnetic resonance brain images and positron emission tomography brain images,and the study included the following aspects: first,the total of 880 patient images obtained from the database were preprocessed,and the processing flow was: format conversion,cranial stripping,bias field correction,and alignment;then,the 3D-DenseNet depth model is applied to 3D medical images for feature extraction,meanwhile,3D-DenseNet-B and3D-DenseNet-BC structures are added for control experiments,including unimodal input,multimodal input for a total of 5 experimental groups,and the features extracted are fused,and finally,the prediction results are obtained after the classifier.The innovation points of this study mainly include:(1)to solve the problem of small brain lesion features and inconspicuous brain area changes in patients with early mild cognitive impairment,3D-DenseNet was selected for the experimental feature extraction work,and the information acquisition rate was improved by capturing detailed information of brain area changes through feature reuse;(2)to improve the network training speed.The study incorporates 3D-DenseNet-B and 3D-DenseNet-BC networks,and reduces the number of feature channels,improves the computing speed of the network and reduces the training time by adding convolutional kernels within the dense blocks of the network and by adding compression parameters in the transport layer;(3)To improve the prediction accuracy.The experiments use the efficient channel attention module as the base architecture to perform feature fusion,which improves the weight of the training model on the target region of focus,highlights the significant useful features,and improves the prediction accuracy.This work,on the one hand,effectively improves the prediction accuracy for patients with mild cognitive impairment and,on the other hand,demonstrates that the combination of neuroimages with different modalities can provide more feature information and help to understand the onset progression of patients with mild cognitive impairment by controlling the results of the experimental group.