Research On Brain Aging Prediction Model Based On Magnetic Resonance Imaging Technology

The brain will inevitably undergo tissue,structure and functional decline during aging.Therefore,the study on the mechanism of brain aging is of great significance for the early prevention,diagnosis and treatment of diseases related to brain aging.An age prediction method based on brain imaging data can predict individual age.By comparing the predicted age with the real age,the deviation of brain structure and function during brain aging can be captured.Based on the deviation,biomarkers related to brain aging can be found.The biomarkers have important implications for the diagnosis of clinical diseases.Therefore,the purposed of this study are two folded.First,to obtain brain images with the help of magnetic resonance imaging technology,and second,to explore the gap between the actual age and the age predicted based on brain images.Previous age prediction studies generally relied only on structural or functional MRI data of the brain.In this study we employ multimodal data fusion to improve the predictive power of the model.The research contents include the following aspects:1.An age prediction study was conducted based on 648 healthy subjects aged between 19 and 88 years old.Firstly,extracting features,including extracting brain volume and cortical index based on T1-weighted image,constructing structural connection matrix and extracting connection weight of matrix based on diffusion tensor image and resting state functional magnetic resonance image construct the function connection matrix and extract the connection weight of the matrix;secondly,reducing feature dimension;finally,the features obtained in the above process are used to construct the age prediction model of different modes.In selecting predictive models,we compared several classical machine learning algorithms,and finally concluded the age prediction model based on Bayesian ridge regression and brain volume feature space.Thus,we can get the optimal prediction results.Based on the results of the Bayesian ridge regression prediction model,we can propose features with significant feature weights as the proposed age-related biomarkers.In this study,it was found that in the characteristic space of average cortical thickness,the features significantly related to age were mainly concentrated in the frontal lobe,temporal lobe,insula and other sulci.In the characteristic space of average cortical surface area,it was found that the features significantly related to age were mainly concentrated in multiple regions such as temporal lobe,anterior and posterior central sulcus.In the feature space of the functional connected network,the features significantly related to age are mainly concentrated in the temporal-frontal lobe and the parietal-frontal lobe.In the characteristic space of structural connection network,the features significantly related to age were found to have more connections in the frontal lobe.Most of the temporal lobe regions were concentrated in the connections between the frontal and temporal lobes,and there are also connections between other cortical brain regions and subcutaneous brain regions.The structural connections in these regions are primarily associated with declines in memory and other cognitive functions.2.The data of each mode not only has similar shared information,but also has its own unique information of modal characteristics.Therefore the age prediction model that integrates multiple modal data may get better prediction results.The brain volume and cortical features obtained from the T1-weighted image,the functional connection matrix obtained from the resting state magnetic resonance imaging data and the structural connection matrix obtained from the diffusion tensor image are taken as the eigenvalues.With the aid of the Stacking of ensemble learning algorithm as the age of the overall architecture model,primary learning with Bayesian ridge regression,the Stacking of the integrated learning algorithm is used as the age prediction of the overall architecture.The model,the primary learner uses Bayesian ridge regression,and the secondary learner uses linear regression as the prediction algorithm and uses random forest regression and Adaboost algorithms for comparison.The data of multiple modalities is found to be the lowest error with the data fusion of function and structure,with an average absolute error of 5.53 years old,the Pearson correlation value r is 0.93.However,the minimum average absolute error of single-mode age prediction is 6.60 years old.Therefore,the age prediction of multimodal fusion is obviously better than the optimal prediction of the age prediction of a single modality.To sum up,in this paper,age prediction model was constructed by fusing structural and functional magnetic resonance data,and significant features of brain images were extracted as age related biomarkers,which can provide some references for clinical diagnosis.