高精度3D光流驱动的DTI空间归一化算法：超越传统与自动化

20 浏览量更新于2024-08-27 收藏 895KB PDF 举报

本文主要探讨了在基于体素的扩散张量图像（DTI）分析中至关重要的空间归一化问题。研究者针对这一挑战，提出了一个基于高精度的3D光流技术的新型DTI数据高维度空间归一化算法。常规的光流理论在处理图像时，强调在保持时空连续性的同时，确保强度一致性及强度梯度一致性。然而，这个新算法突破了传统限制，采用了分级策略，从粗粒度到精细尺度，通过Euler-Lagrange数值分析方法，实现对DTI数据的精准注册。在传统光学流算法和简单的仿射对齐方法之外，该算法展示了显著的优势。它能够在保持图像结构完整性的同时，提高空间归一化的精确度。实验证明，无论是通过模拟数据还是实际临床数据，相比于统计参数映射（SPM）等常用工具，该算法展现出更高的准确性。这在DTI分析中尤其重要，因为它能确保数据的准确比较和统计分析。此外，该算法的一大亮点是其高度自动化。它需要的参数极少，几乎不需要人工干预，极大地节省了时间和人力资源。这对于大规模DTI数据处理和研究具有重大意义，使得研究人员可以更加专注于数据分析本身，而无需花费过多精力在预处理步骤上。这项研究对于提高DTI分析的标准化和精确性具有革命性的影响，不仅提升了空间归一化的性能，还简化了工作流程，有望推动神经影像学领域内的研究进展。在未来的研究和应用中，这种基于光流的高精度空间归一化算法可能会成为DTI数据分析的标准方法之一。



Abstract

—Spatial normalization plays a key role in

voxel-based analyses of diffusion tensor images (DTI). We

propose a highly accurate algorithm for high-dimensional spatial

normalization of DTI data based on the technique of 3D optical

flow. The theory of conventional optic flow assumes consistency

of intensity and consistency of the gradient of intensity under a

constraint of discontinuity-preserving spatio-temporal

smoothness. By employing a hierarchical strategy ranging from

coarse to fine scales of resolution and a method of

Euler-Lagrange numerical analysis, our algorithm is capable of

registering DTI data. Experiments using both simulated and real

datasets demonstrated that the accuracy of our algorithm is

better not only than that of those traditional optical flow

algorithms or using affine alignment, but also better than the

results using popular tools such as the statistical parametric

mapping (SPM) software package. Moreover, our registration

algorithm is fully automated, requiring a very limited number of

parameters and no manual intervention.

NTRODUCTION

any methods in image analysis have been developed to

tackle DTI registration based on the strategy of how to

set up the correspondence between two subjects, a

strategy that generally falls into one of two categories. The

first involves feature-based matching, i.e. transformations that

are calculated based on a number of anatomical

correspondences established manually, semi-automatically, or

fully automatically on a number of distinct anatomical

features, including distinct landmarks [1,2] or a combination

of curves and surfaces, such as sulci and gyri [3,4,5]. In

addition, finding anatomical correspondences is a key to the

success of the spatial normalization. To find anatomical

correspondences, conventional methods generally first extract

scalar features from each tensor individually; and then by

constructing scalar maps, regional integration and other

operations such as edge detection, a final set of features for

correspondence matching can be constituted. For example,

Manuscript received March 1, 2013. This work was supported by The

National Natural Science Foundation of China (No. 61273261 & 91232701),

Shanghai Pujiang Program (No. 12PJ1402800), Shanghai Knowledge

Service Platform Project (No. ZF1213) and also was partly supported by

Shanghai Commission of Science and Technology (No. 10440710200) .

Ying Wen is with the department of computer science and technology,

East China Normal University, Shanghai, 200241, China (e-mail: ywen@

cs.ecnu.edu.cn).

Bradley S. Peterson is with the Brain Imaging Laboratory/MRI Unit,

Department of Psychiatry, Columbia University, New York, NY, 10032,

USA (e-mail: petersob@nyspi.columbia.edu).

Dongrong Xu is with the Brain Imaging Laboratory/MRI Unit,

Department of Psychiatry, Columbia University, New York, NY, 10032,

USA (phone:2125435495; fax:2125430522; e-mail:dx2013@columbia.edu)

Timer [6] and F-Timer [7] were developed based on Hammer

[8], which adopts so-called driving voxels in addition to

geometric moments as the features upon which registration is

based. The boundary-based registration uses the measure of

the thickness of cortex in T1-weighted image [9]; Yang et al.

guided their registration using information on a tensor’s

structural geometry and orientation [10]; and Xue et al.

proposed a local fast marching method for DTI registration

[11]. These landmark-based methods numerically seek

optimal transformations (under respective model-specific

assumptions) to maximize similarity across brains.

The second category of registration methods is based on

volumetric transformations between one brain image and a

template image. These methods assume that the images are

acquired using the same pulse sequence [12,13,14] and these

methods therefore are generally more efficient, because they

do not require the construction of a specific anatomical model

using anatomic landmarks. Christensen et al [15] proposed a

viscous fluid flow model to implement registration. Unlike

fluid flow, optical flow is another volumetric transformation

method. The technique of optical flow developed in computer

vision for object tracking [16,17] is one such method. The two

approaches differ with the constraints in their

implementations. The former adopts viscous constraints and

body force while the latter adopts gradient and smooth

constants and Taylor expansion. Most of the prevailing

approaches for DTI normalization, for example, the

preservation of principal direction (PPD) [18], the

Procrustean normalization method [19,20], actually make use

of the registration method for conventional scalar images to

calculate the voxel correspondence across two imaging spaces,

following which an additional step is required to properly

reorient the tensors. Although these methods work well, they

do not yet fully consider the characteristics of DTI data, and

the 2-step schemas are computationally inefficient. Therefore,

a method for spatial normalization of DTI data that takes

advantage of optic flow would be highly attractive to the

neuroimaging community.

We propose in this work a novel approach for DTI

normalization using optical flow. We first extend the

traditional 2D optic flow to a 3D form, and then incorporate it

into a hierarchical framework that is applied to imaging data

in multiple resolution scales from coarse to fine, thereby

building up a registration approach with high accuracy.

II.

ATERIAL AND

ETHODS

Data Acquisition

We acquired all the human Diffusion-Weighted Imaging

A Highly Accurate, Optical Flow -Based Algorithm for Nonlinear

Spatial Normalization of Diffusion Tensor Images

Ying Wen, Bradley S. Peterson, Dongrong Xu

下载后可阅读完整内容，剩余7页未读，立即下载

weixin_38549721

粉丝: 9
资源: 882

高精度3D光流驱动的DTI空间归一化算法：超越传统与自动化

基于变分理论的高精度光流估计（brox光流经典）

matlab开发-GLMDTPSRegistrationMethod

MIBCA工具箱的使用说明书（详细）

用均值和标准差归一化张量图像

张量归一化python

请用matlab实现基于梯度扩散和图像的结构张量对图像进行降噪处理，在达到降噪的同时并最大化地进行保边

针对tensor张量计算Min-Max 归一化

在DTI Studio 2.30中，如何利用彩色映射功能增强扩散张量图像的3D可视化效果，并调整哪些参数来实现最佳的视觉效果？

pytorch 归一化张量

张量怎么归一化处理详细说明

最新资源