Research On Some Problems Of Infant Brain Development Based On Neuroimaging Computing And Machine Learning

The infant brain undergoes active development in both structure,function and cognition during the first postnatal years.Exploring the development status of infant brain in this postnatal duration presents great importance for the neuroscience research field,e.g.,help to understand the early development pattern of the healthy infant brain,and help to early diagnose the neurodevelopment disorders or psychiatric diseases of the infant brain.Recently,the rapid developments of magnetic resonance imaging,neuroimaging computing,and computer science technologies have provided great convenience for accurately mapping the infant brain structure and studying the infant brain development.In this thesis,based on neuroimaging computing and machine learning,a large longitudinal magnetic resonance image dataset of infant brain(>900 infants)is used to study three important aspects of infant brain development through the features of infant cortical folding:1)exploration of the major cortical folding patterns of infant brain based on multi-view curvatures;2)infant identification and individual variability study;3)prediction of infant cognitive scores based on a spherical neural convolutional network.Exploring the major cortical folding patterns of infant brain based on multi-view curvatures and unsupervised learning.The cortical folding patterns of infant brain have significant individual differences.In the large-scale dataset,for some cortical regions with high individual variability,the morphologies of their cortical foldings across different subjects are variable,might present several representative cortical folding patterns.Exploring the cortical folding patterns of different cortices will help us to understand the brain morphology of healthy infants;at the same time,since some neurodevelopmental disorders would lead to morphological changes in the cerebral cortex,understanding the normal morphology of cortical folding pattern will also help to detect the abnormal morphology;besides,based on the obtained representative cortical folding patterns,multi-atlas can be constructed,which will improve the accuracy of cortical surface registration compared to a averaged single atlas.This study proposes a novel unsupervised learning method to explore the major cortical folding patterns present in cortical regions with high individual variability.Specifically,we first reconstruct the cortical surface with spherical topology in the manifold space based on the magnetic resonance image,and extract the multi-view curvatures based on the mean curvature feature and the spherical wavelet transform.Then,the similarity matrix is constructed for each view of feature,and the similarity matrices of all features are nonlinearly fused into a comprehensive similarity matrix by the similarity network fusion method.The matrix contains both shared and complementary similarity information from different views of features across all the subjects.Finally,based on the fused similarity matrix,all the subjects are clustered into different groups using the hierarchical affinity propagation algorithm,thus obtaining the typical cortical folding patterns for each group.We apply the proposed method to the superior temporal gyrus(STG),inferior frontal gyrus(IFG),cingulate cortex,precuneus,and explore the major cortical folding patterns of these regions.In addition,we find that the cortical folding patterns present gender differences in STG,IFG,and cingulate gyrus,and present hemispherical asymmetry in STG and cingulate cortex.We also apply the proposed method to the human connectome project(HCP)dataset,and compare the results with the results of infants.The results indicate that the maj or cortical folding patterns of the adult brain already present at birth.Infant identification and individual variability study based on the multi-scale regional folding descriptor.The infant cerebral cortex is under dynamic postnatal development in the first years of life.Infant identification based on the cortical folding patterns of neonates is of great importance in two aspects:on the one hand,it helps to explore whether the individual variability in the early developmental cerebral cortex is stable and reliable for infant identification;on the other hand,it helps to explore which cortices express more individual variability and thus contributing more in the infant identification tasks;in addition,based on the twin samples,it can help investigate whether the twins,especially identical twins,can be accurately identified by the cortical folding patterns.This paper proposes a novel regional descriptor,i.e.,FoldingPrint,based on multi-scale decomposed curvatures to characterize the cortical folding of brain regions,and proposes a simple but efficient method for infant identification based on the similarity matrix and voting scheme.We applied the proposed method to a longitudinal infant dataset,using the scans at birth to identify their corresponding scans at year 1 and year 2,respectively.The experimental results show that the infant can be accurately identified based on the neonatal brain with the proposed FoldingPrint descriptor.Besides,the cortical regions with high identification ability largely overlap with the cortical regions with high individual variability.In addition,the proposed FoldingPrint can be effectively used for the individual identification of twins.Even for the identical twins with the same gene and similar developmental environments,they still can be accurately identified based on the FoldingPrint.Prediction of infant cognitive scores based on spherical convolutional neural networks.The structural features and cognitive function undergo dynamic developmental changes during infancy.Cognitive score prediction is important for understanding the cognitive developmental status of healthy infants and for further understanding the infant cognitive impairment and its early diagnosis.However,little is known about whether the brain structure is related to cognitive function,and whether the structural features can be effectively used for predicting the cognitive scores.Using the inherent spherical topology of the cortical surface in manifold space,we propose a deep learning method based on spherical convolutional neural network,and predict the infant cognitive scores with cortical features.Specifically,we transfer the convolution operation in the Euclidean space to the manifold space,thus constructing the spherical convolutional filter.Then,we define the corresponding spherical convolution and spherical pooling in the triangle mesh structure.Next,these spherical operations are embedded in the DenseNet block to replace the original convolution and pooling operations,thus constructing the Spherical-DenseNet.This deep network structure can be used for end-to-end prediction for five standard cognitive scores(fine motion,visual reception,language expression,language reception,early learning composite).The proposed method is applied in a large longitudinal infant brain MRI dataset,in which each subject was scanned at birth,1 and 2 years of age,and each subject accepted a cognitive ability test at 2 years of age and obtained cognitive mullen scores.The cortical features of infant brain at three-time points are combined to predict the mullen scores of this specific infant.The experimental results indicate that the proposed Spherical-DenseNet outperforms the traditional machine learning and deep learning methods,and obtains better prediction performance.