Prediction Of Brain Maturation Using Multimodal MRI Data Across Human Lifespan

During the development of the brain,its structure and function have changed significantly.Brain age prediction based on magnetic resonance imaging(Magnetic Resonance Imaging,MRI)provides a lot of useful information for evaluating brain development level,disease prevention or slowing down of brain disease,and clinical diagnosis and treatment.This study uses multimodal MRI and machine learning methods to predict the age of the brain.It aims to evaluate the predictive performance of multimodal MRI-related features,find brain regions that have an important contribution to age prediction,and explore some factors that affect age prediction.The content of this study is as follows:Firstly,Brain age prediction based on Support Vector Regression(SVR)and Gaussian Process Regression(GPR)algorithms.For multimodal MRI data containing brain structural information and functional information(with the open dataset: Nathan Kline Institute-Rockland Sample data set,the dataset is 505 samples of all ages from 6 to 85 years old).We use GPR and SVR to evaluate the age prediction performance of seven MRI features and the combination of seven features: low frequency amplitude(ALFF),fractional low frequency amplitude(fALFF),regional homogeneity(ReHo),cortical surface area,cortical thickness,mean curvature,gray matter volume and combination of seven features.The research results show that,in terms of the seven single-modality features and their combined features,we found that the prediction performance of gray matter volume and cortical thickness is relatively good among all the individual features;in fMRI features,the prediction effect of ALFF is better than ReHo.The feature prediction performance of the combination for four T1 weighted features and the three functional features is better than the prediction performance of any single modality.In addition,it is also found that ReHo for the prediction performance,play a negative role.For the T1 weighted features,the salient brain regions are Subcentral gyrus and sulci,Middle temporal gyrus,Opercular part of the inferior frontal gyrus,Superior segment of the circular sulcus of the insula,Subcentral gyrus and sulci,Postcentral gyrus,cingulate gyrus,Middle frontal gyrus,Superior frontal gyrus,Middle occipital gyrus,Superior occipital gyrus and so on.For fMRI features,the map shows that the frontal lobe,the corpus callosum and other brain areas.Secondly,Brain age prediction based on Partial Least Square Regression(PLSR)algorithms.Because GPR and SVR operations are relatively complex,time-consuming,and have a lot of room for improvement in accuracy,we use the PLSR algorithm with lower complexity and can deal with the multiple collinear problems between the indicator features to evaluate the age prediction performance of seven MRI features,and through weight coefficients to find brain regions that contribute significantly to age prediction,and explore factors that affect age prediction.The research results show that the order of features with high to low prediction capabilities is: cortical thickness,the combination of all seven features,the combination of four features in the T1 weighted MRI,fALFF,grey matter volume,a combination of three features in the fMRI,ReHo,mean curvature,ALFF,and surface area.Moreover,we found that some specific brain regions have a significant contribution to brain age prediction.The regions are Temporal pole,the circular sulcus of the insula,Central sulcus and gyrus,Inferior occipital gyrus and sulcus,Precuneus,Inferior temporal gyrus,Lateral orbital sulcus,Subcallosal area,Lateral aspect of the superior temporal gyrus,Calcarine sulcus and so on.This is roughly the same as the results based on the SVR and GPR algorithms,but some new brain areas were Calcarine sulcus,Precuneus and Lateral orbital sulcus.In addition,we also found that the prediction accuracy increases with the length of the time point,but the increase rate is relatively small.And,the prediction accuracy of fMRI features increased significantly with the number of subjects,up to about 200.When the number of subjects is greater than 200,the increase rate is relatively small.For T1 image features,the prediction accuracy remains fairly consistency in the range of 78 to 505 subjects.Therefore,we recommend that you not need a lot of data to make an age prediction.In a word,the main findings of our research indicate that certain brain regions play a crucial role in the prediction of brain age.Moreover,multi-modal features help to improve the accuracy of age prediction,but these results do not indicate that multiple modes must be better than a single mode,nor does it mean that a large amount of MRI data can lead to better analysis.