Scalable Image Registration and 3D Reconstruction at Microscopic Resolution - High-Throughput Image Reconstruction and Analysis - page 185

tion of expression patterns of genes in mouse embryo. In [10], the authors built
3D models for human cervical cancer samples using stacks of histological images
in clinical settings. A complete study on registering large microscopic images of
mouse brain sections was presented in [22].
8.2.3 HPC Solutions for Image Registration
Large scale image registration has many applications in both biomedical research
[10, 22, 28] and geophysics [30]. However, there are currently few works address-
ing image registration algorithms intended to run efficiently on HPC environments.
The work on parallel image registration on multicomputers is limited [22]
and is restricted to either large computer clusters [31--33] or IBM cell clusters
[34]. Clusters of GPUs have been used to implement other heavy workload tasks
[35], mostly within the simulation and visualization fields. For example, numerical
methods for finite element computations used in 3D interactive simulations [36],
and nuclear, gas dispersion, and heat shimmering simulations [37].
On the other hand, commodity graphics hardware has become a cost-effective
parallel platform to implement biomedical applications in general [38]. Applica-
tions similar to ours such as the registration of small radiological images [39] and
computation of joint histogram for image registration [40], and others within the
fields of data mining [41], image segmentation, and clustering [42] have applied
commodity graphics hardware solutions. Those efforts have reported performance
gains of more than 10 times [43], but they were mostly implemented using shaders
with the Cg language [43].
8.3 Two-Stage Scalable Registration Pipeline
To address the specific challenges of nonrigid distortion, large image size, and
feature rich content, we have proposed an algorithm that consists of two stages:
rigid initialization and nonrigid registration. Rigid initialization estimates the rough
alignment of the base-float image pair from the consensus of correspondences be-
tween anatomical features. The nonrigid stage seeks to establish pixel-precision
correspondences by precisely matching areas of intensity information. The initial-
ization reduces the search for matching in the nonrigid stage, resulting in a lower
likelihood of erroneous matches and less computation. This combination provides
an approach to the problem of automatic sectioned image registration and recon-
struction that is robust, easily parallelizable, and scalable.
8.3.1 Fast Rigid Initialization
The basis of the rigid initialization stage is the matching of high-level features or
image regions that correspond to anatomically significant features such as blood
vessels, mammary ducts, or small voids within the tissue boundary. This is a natu-
ral choice for features that has several advantages over the more primitive features
generated by commonly used methods such as corner detection. First, the amount
of high level features is relatively limited keeping the total number of possible
matches reasonable. Second, the descriptions used to match these features such
as shape, size, and type are invariant under rotation and translation and so the