Brain Age Prediction Based On Structural Magnetic Resonance Imaging

The human brain is constantly changing in the whole human life cycle,and these changes partly reflect the brain’s normal aging process.However,the age of the brain may not be equal to the actual physical age because of the factors such as abnormal brain development,neuropsychiatric disorders,or brain diseases.Efficient brain age prediction methods can not only help to deepen our understanding of the process of brain aging,but also be conducive to objective diagnosis and early warning of neuropsychiatric disorders.There will be obvious changes in brain structure during the development and aging of the human brain,and structural magnetic resonance imaging technology can provide high-resolution three-dimensional images of human brain tissue,Therefore,the human brain structure magnetic resonance images have become important data bases for brain age prediction studies.At present,a large number of brain age prediction researches based on structural magnetic resonance images have been carried out,but there are still some common problems: Most brain age prediction studies based on 3D-CNN take the whole-brain image as input,accompanying high model complexity;Most of these studies are based on unbalanced data sets,with fewer sample sizes for children and the elderly than middle-aged adults,and the imbalance of samples is generally not considered;These researches are generally based on single center data to extract single template features,without fully considering the differences of different central data sets and the complementarity of information provided by different templates.To address the issues,we performed three sets of analyses as follows:(1)Ensemble 3D-CNN brain age prediction based on multiple regions of interest.The current widespread brain age predictions based on 3D-CNN mostly take images of the entire human brain as model input.The high-dimensional characteristics of brain image mean high computational complexity of brain age prediction based on 3D-CNN.Thus,this study proposed ensemble 3D-CNN based on multiple regions of interest.Instead of putting the entire human brain image into the network,the strategy selects several regions of interest that are more relevant to brain age from magnetic resonance images and trains multiple 3D-CNN networks.The final prediction is the integration of multiple network predictions.The results show that the individual brain age prediction of ensemble 3D-CNN based on regions of interest effectively reduces the time spent on model training,and the prediction accuracy(the correlation coefficient of prediction and real value R=0.873)is improved compared with the classical 3D-CNN(R=0.739).(2)Brain age prediction based on adaptive sample weighting strategy.The data set used in this study showed a spindle-shaped distribution with fewer at both ends and more in the middle(there are fewer samples for children and the elderly and more for middle-aged adults),and the imbalance of the samples will have a negative impact on the prediction accuracy of the model.To address the issue,this study introduces an adaptive sample weighting strategy,which takes into account the proportion of the approximate age interval of the sample to the whole data set when calculating the loss: a larger weight is given to increase the penalty cost when the sample is located in the age range with fewer sample size.The results show that the brain age prediction based on adaptive sample weighting strategy(R=0.883)can effectively reduce the age prediction error for children and the elderly.(3)Brain age prediction based on multi-center multi-template.At present,most brain age prediction models use data from a single imaging center,and utilize only a single brain template to extract features.However,data sets in a single imaging center are insufficient,and using only a single template can easily cause the problem of insufficient feature information.Therefore,this study proposed a multi-center multi-template ensemble prediction method.The method integrates multiple center data sets and feature set information of multiple templates to fully utilize multi-center data and fully explore the effective complementary features provided by multiple templates.The results show that the brain age prediction based on multi-center multi-template(NKI data set: R=0.702,ADNI data set: R=0.726)is better than that of single-center single-template(NKI data set: R=0.641,ADNI data set: R=0.659).The innovation of this study are: 1)introducing ensemble 3D-CNN based on multiple regions of interest to solve the problem of high complexity and overfitting of the classical 3D-CNN model,which effectively improves the efficiency of model training and the accuracy of brain age prediction;2)An adaptive sample weighting strategy is introduced to solve the problem of unbalanced sample distribution,which effectively reduces the age prediction error for children and the elderly that sample size of these people are small;3)The multi-center multi-template strategy effectively utilizes the multi-center data and multi-template feature and improves the prediction accuracy of brain age.