Research On Digital Statistical Atlas Of Three-dimensional Human Brain Diffusion Tensor Imaging

Diffusion tensor imaging technique is adopted to capture the intensity, direction and anisotropy of water diffusion, which is convinced to reflect the micro-structure of biological tissues. As a non-invasive imaging technique, brain DTI images can provide more information about white matter fiber traveling patterns and structures than traditional CT and MRI images. Benefiting from this advantage, DTI has largely broadened the framework of biomedical image collection, pre-and post-processing, and statistical analysis in human brain connectome research and brain diseases diagnosis. Moreover, DTI has contributed to develop new representation theory of brain micro anatomical features and functional characteristics, which has drawn more and more attention in its research. Digital statistical atlas is another important topic in computational neuroanatomy. Construction of statistical atlas based on a large number of normal subjects’medical images can give an abstract description of normal brain anatomy, and the procedure of inter-subject learning and anatomy modeling yields generalization ability, which is crucial to population-based morphometry variation tests.It’s very important to study the construction of three-dimensional DTI statistical atlas in detail. Research in this field has just started, both in domestic and abroad. Essentially, a3D DTI image is equivalent to3D second-order tensor field. The construction of DTI atlas relates to tensor estimation and optimization, tensor field registration, feature representation in tensor field, and atlas estimation. In this paper, techniques of digital statistical atlas construction for scalar medical images, like CT and MRI, have been discussed in detail. Furthermore, these techniques were modified and extended to tensor field for DTI images. Finally, the DTI atlas formed in this paper was validated and evaluated within the application of human brain connectome modeling. The major contributions of this dissertation are as follows.1. Riemannian manifold and Log-Euclidean metrics were introduced to improve the basic tensor model and the calculus theory of tensor field was reconstructed, which provided the fundamentals of registration, feature representation in tensor field. Meanwhile, the tensor interpolation and regularization algorithm were improved by using Log-Euclidean metrics. Experiments illustrated the improvement of tensor estimation and fiber tracking. 2. An improved diffeomorphic image registration algorithm in tensor field was put forward based on Log-Euclidean metrics and tensor reorientation. Firstly, LDDMM （large deformation diffeomorphic metric mapping） and diffeomorphic Demons were studied and extended to3D2nd-order tensor field. Then, by incorporating finite strain theory, tensor reorientation was added to diffeomorphic Demons, which was proved to keep the intrinsic geometric features of tensor during the deformation procedure. Experiments showed that diffeomorphic Demons overwhelmed with less computational burden, which was more applicable for atlas construction in vast population.3. On the basis of previous research in high-resolution brain CT texture extraction, a novel multi-scale feature extraction method, named TIDA （tensor image discriminative attributes）, was put forward. By integrating the tensor geometric parameters and boundary information, TIDA revealed considerable discriminative power in segmenting brain tissues, like white matter, gray matter and cerebrospinal fluid, as well as brain occupying lesions.4. An automated DTI digital statistical atlas construction framework was put forward. In this framework, Log-Euclidean diffeomorphic Demons was performed to register inter-subject DTI images, and TIDA provided the multi-scale representation of DTI image features. The most important thing was that the diffeomorphic and non-biasd property had been guaranteed by the improved atlas construction procedure.5. Within the open-source research database of Human Connectome Project,20cases of normal brain3D DTI images were used to create statistical atlas, named Beijing_Zang₂0. And white matter fiber bundles were tracked in the intact brain DTI atlas, with respect to the brain parcellation. The network anatomy of DTI atlas was studied in by incorporating MNI152and ICBM-DTI-81atlas. Various types of network characteristics were generated accordingly.