Prediction Of Age Based On Multimodal Magnetic Resonance Imaging

The research on the maturation and aging of the human brain based on magnetic resonance images has been a hotspot in the field of neuroscience in recent years.Previous studies demonstrated the structural and functional changes of the brain with development and aging,and these results indicated the potential of magnetic resonance images,in combinations with modern machine learning techniques,in the prediction of individuals’ age.Researches also showed obvious deviations between the brain age and his/her chronological age in patients with neuropsychiatric disorders.Therefore,the prediction of individual brain age will not only help assess the degree of brain development/aging,but also provide additional information for the diagnosis of neuropsychiatric disorders,such as attention deficit hyperactivity disorder(ADHD)and Alzheimer’s disease.In this study,the prediction of individuals’ brain age was carried out by using the semi-supervised co-training regression(COREG),E-net and random forest algorithms based on the features extracted from multimodal magnetic resonance imaging data.The details are as follows:(1)Prediction of brain ages based on the features describing the static brain networks.By combining two types of features,the resting-state functional connections(RSFC)extracted from the resting-state fMRI and the FA parameter extracted from the DTI data of 73 adult subjects aged 30-85 years old,we predicted individuals’ brain ages with the method of semi-supervised co-training regression algorithm(COREG).The correlation between the predicted and real age for the best prediction was 0.78(MAE ＝8.39 years),and this indicates that it is effective to predict individuals’ brain ages based on the features describing static brain network.Furthermore,the contribution of FA is greater than that of RSFC among the optimal prediction models,which indicates that FA parameter is more sensitive to the aging of the human brain than RSFC.(2)Prediction of brain age based on the features describing the dynamic brain networks.We predicted individuals’ brain ages with the use of elastic net(E-net)and random forest(RF)respectively,based on two sets of features,namely,the variability(Mean)and amplitude(Mean-Abs)of FC dynamic fluctuations extracted from the resting-state fMRI data of 117 adult subjects aged 30-85 years old.The results showed that the predicted ages based on the Mean of FC dynamic fluctuations(E-net:R=0.53,MAE=13.07 years;RF:R=0.58,MAE=12.62 yesrs)was much better than those based on the Mean-Abs of FC dynamic fluctuations(E-net:R=0.46,MAE=13.87;RF:R=0.56,MAE=13.32).In general,the brain age predictions based on the features describing the static brain networks are better than those based on the features describing the dynamic brain network.One possible explantion is that the features describing the dynamic brain networks may include more random noises.(3)Prediction of brain ages based on the structural features of the human brain.We predicted individuals’ brain ages with the use of RF based on the density of gray matter extracted from the sMRI data of 706 subjects aged 7-22 years old.The results showed that the prediction of individuals’ brain ages based on the density of gray matter perform the best:the correlation between the predicted and real age was 0.82(MAE=1.59 years).This result indicates that the density of gray matter is more sensitive to brain development,and may be a more effective feature for brain age predictions.Overall,in this study,the prediction of adult,teenage and puerile brain age was carried out by using the method of machine leaning,such as semi-supervised co-training regression(COREG),E-net and random forest algorithm respectively,based on the features extracted from the DTI,fMRI,and MRI data.The results of this study suggest that the structural features of the human brain(such as FA and density of gray matter)are more effective for brain age estimation as compared to functional features(such as RSFC and dynamic function connectivity).The innovations of this study are as follow:1)co-training was used in this study for feature fusion;2)the features describing the dynamic network of the human brain was used to predict individuals’ brain ages.