Scalable Image Registration and 3D Reconstruction at Microscopic Resolution - High-Throughput Image Reconstruction and Analysis - page 183

address computing requirements, implementation is presented in several HPC
environments including CPU clusters and graphical processing units (GPU).
Additionally, a scheme is proposed for preserving the integrity of 3D duct-like
structures in tissue reconstruction. This work is motivated by several large scale
biomedical studies including developmental biology and breast cancer research.
This chapter is organized in the following sections. In Section 8.2, existing ap-
proaches for image registration are reviewed with emphasis on several large scale
research projects that require the alignment of 2D microscopic slides for 3D re-
constructions. Existing HPC solutions related to image registration and 3D recon-
struction of microscopic structures are also discussed. In Section 8.3, the two-stage
registration algorithm is presented along with a discussion on the preservation of
3D ductal structures in reconstruction applications. Section 8.4 discusses HPC so-
lutions for image registration based on the two-stage algorithm, including parallel
computing using CPU and GPU clusters. In order to evaluate the implementations
described in the previous section, the two stage registration algorithm was applied
in a high performance computing environment to a benchmark of images taken
from biological studies in Section 8.5. The results for the algorithms and HPC
implementation are presented in Section 8.6.

8.2 Review of Large-Scale Image Registration

8.2.1 Common Approaches for Image Registration
Image registration has been extensively studied in many applications including radi-
ology, geological survey, and computer vision. It can be framed as an optimization
problem; that is, finding the transformation T between two images I 1 and I 2 to
maximize their similarity with respect to some measure,

T = arg max Similarity ( I 1 , T ( I 2 )) (8.1)

In this context, a registration algorithm needs to specify the following condi-
tions:

1. The similarity metric which determines the cost function for the optimiza-
tion process. The commonly used similarity metrics include mutual infor-
mation (MI) [23], normalized cross-correlation (NCC), and summed square
difference (SSD) [7, 8]. Among them, MI was originally designed for registering
images of different modalities such as magnetic resonance image (MRI) with
positron emission tomography image (PET) for human brain scanning. MRI
images provide structural information while the PET images supply functional
information. By maximizing the mutual information between the two types of
images, the user obtains an overlay of the functional map on a structural atlas
of the brain. NCC and SSD are frequently used for matching images of the
same modality. In our work, we use NCC given that we are dealing with histo-
logical images with the same type of staining. However, we also observed that
for histological images with different types of staining, a variation of NCC can
still be used [28] due to similar anatomical structure between the histological