Scalable Image Registration and 3D Reconstruction at Microscopic Resolution - High-Throughput Image Reconstruction and Analysis - page 181

CHAPTER 8
Scalable Image Registration and
3D Reconstruction at Microscopic
Resolution
Lee Cooper, Kun Huang, Manuel Ujaldon, and Antonio Ruiz
8.1 Introduction
Characterizing the phenotypes associated with specific genotypes is critical
for understanding the roles of genes and gene interactions. Three-dimensional
morphologies of cellular and tissue structure are one aspect of phenotype that
provides information key to the study of biological processes such as the initiation
of cancer in the tumor microenvironment, the development of organs via complex
induction processes, or the formation of neural networks in the brain cortex.
Nevertheless, existing techniques for obtaining high-magnification 3D information
from biomedical samples are rather limited. While confocal and multiphoton
imaging offer 3D capability, both are limited in field and depth and only structures
with fluorescent labeling are visible. Therefore, another fundamental approach
for 3D acquisition is to perform reconstruction from multiple 2D images obtained
from tissue sectioning. The basis for this approach is the process of automatically
aligning images of serial sections using image registration [1--22] to compensate
for misalignments and distortions.
Image registration has been extensively studied in biomedical imaging,
geological survey, and computer vision [7, 8]. It can be considered as an
optimization problem for finding the optimal transformation T between two
images I 1 and I 2 to maximize their similarity. Commonly used similarity metrics
include mutual information [23], cross-correlation, and summed square difference.
The transformation spaces include rigid transformation, which deals with only
rotation and translation, and nonrigid transformation which compensates for
deformations such as bending, stretching, shearing, and warping [11, 24, 25].
Like any optimization process a good initialization is critical for a global optimum
outcome. In many cases, a good rigid registration result serves as an ideal
initialization for nonrigid registration [10]. For large images with conspicuous
deformations, hierarchical multiresolution registration methods have also been
widely used in medical imaging applications [26, 27].
There are four primary issues for the registration of section images for tissue
reconstruction:
1. Heavy computational requirements. High-resolution slide scanners are
capable of generating images with resolutions of 0.46mm/pixel (with 20X
objective lens), often producing images with hundreds of millions or
181