Spherical Neural Network Based Analysis Of Cortical Surface Data

The advance of magnetic resonance imaging(MRI)technology makes it possible to acquire and observe human brain images,and thus study the neurological development and diseases in a non-invasive way.Over the past few decades,MR image processing tech-niques have been developed and matured,in which the most critical steps are brain volume-based analysis and cortical surface-based analysis.Recently,many deep learning-based al-gorithms have been developed to improve the speed and accuracy of volume-based analysis,but surface-based analysis still relies on the traditional hand crafted features and machine learning algorithms,which can no longer meet the requirements of current large-scale neu-roimaging studies.Therefore,to obtain faster and more accurate results of surface-based analysis,in this thesis,we present an comprehensive study on the use of Convolutional Neural Networks(CNNs)for analyzing cortical surface data.Specifically,we focus on the following four subjects.(1)Spherical U-Net-based cortical surface parcellation and development prediction.As the structure of cortical surface has a spherical topology in a manifold space,there is no consistent neighborhood definition and thus no straightforward convolution/transposed convolution operations for cortical surface data.In this thesis,by leveraging the regular and consistent geometric structure of the resampled spherical cortical surface,we propose a novel convolution filter,called 1-ring filter.Accordingly,we develop corresponding oper-ations for convolution,pooling,and transposed convolution for spherical surface data and thus construct spherical U-Net architecture.We then apply Spherical U-Net to two important cortical surface analysis tasks:parcellation and development prediction.Both applications demonstrate the competitive performance in the accuracy and computational efficiency of our Spherical U-Net,in comparison with the state-of-the-art methods.(2)Superfast Spherical Surface Registration(S3Reg)based on three orthogonal Spher-ical U-Nets.Cortical surface registration aligns cortical surfaces across individuals and time points to establish cross-sectional and longitudinal cortical correspondences to facilitate neu-roimaging studies.In this thesis,we develop the S3Reg framework by leveraging an end-to-end unsupervised learning strategy.It offers great flexibility in the choice of input feature sets and output similarity measures,and meanwhile reduces the registration time signifi-cantly.To handle the polar-distortion issue,we construct the three orthogonal Spherical U-Nets architecture to directly learn the velocity fields on the sphere,and then use 6”scal-ing and squaring”layers to guarantee topology-preserving deformations.Experiments are performed to align both adult and infant multimodal cortical features.Results demonstrate our S3Reg shows superior performance with state-of-the-art methods,while improving the registration time from 1 min to 10 sec.(3)Joint cortical surface registration and parcellation based on deep spherical neural network.Conventionally,cortical surface registration and parcellation are performed inde-pendently as two tasks,ignoring the inherent connections of them.To this end,we propose a deep learning framework for joint cortical surface registration and parcellation.Our ap-proach first uses a shared encoder to learn shared features for both tasks.Then we train two task-specific decoders for registration and parcellation,respectively.We further exploit the more explicit connection between them by incorporating the novel parcellation map simi-larity loss to enforce the region-of-interests'boundary consistency,thereby providing extra supervision for the registration task.Conversely,warping one surface with manual parcel-lation map to another surface provides a large amount of augmented data for parcellation task.Experiments on a dataset with more than 600 cortical surfaces show that our approach achieves large improvements over separately trained networks and enables training high-quality parcellation and registration models using much fewer labeled data.(4)Harmonization of cortical property data based on Spherical U-Net.A joint analysis of cortical properties(e.g.,cortical thickness)of multi-site neuroimaging data is unavoid-ably facing the problem of differences in MRI scanners.To address this issue,in this thesis,we combine Spherical U-Net and Cycle GAN to construct a surface-to-surface Cycle GAN(S2SGAN)for harmonizing cortical thickness maps between different scanners.Specifi-cally,we model the harmonization from scanner X to scanner Y as a surface-to-surface translation task.The first goal of harmonization is to learn a mapping G_X:X?Y such that the distribution of thickness maps from G_X(X)is indistinguishable from Y.With the second goal of harmonization to preserve individual differences,we utilize the inverse map-ping G_Y:Y?X and the cycle consistency loss to enforce G_Y(G_X(X))?X(and vice versa).Quantitative evaluation on both synthesized and real cortical data demonstrates the superior ability of our method in removing unwanted scanner effects and preserving indi-vidual differences simultaneously,compared to the state-of-the-art methods.