Scalable Image Registration and 3D Reconstruction at Microscopic Resolution - High-Throughput Image Reconstruction and Analysis - page 182

even billions of pixels. The reconstruction of an individual tissue sample
may involve hundreds of slides, and a full study may contains several sam-
ples with image data easily ranging in the terabytes. Commonly a multiscale
approach is taken to registering images of this size; however, this requires
transformation of the free-floating image at each scale which is computa-
tionally nontrivial. For instance, in this paper we are dealing with images
with sizes up to 23K 62K pixels. With a scale factor of two, this implies

transformation of an image containing around 12K

×

× 31K pixels prior to

the final stage.
2. Feature-rich environment. The textural quality of microscopic image con-
tent provides a unique challenge to the problems of feature selection and
matching. When applied to microscopic image content, traditional feature
detection schemes such as corner detection generate an overwhelming abun-
dance of features that are regular in appearance making matching prone
to error. In addition, at submicron resolutions a shift of 1 mm corres-
ponds to thousands of pixels. The search for corresponding features is
therefore infeasible without good initialization.
3. Nonrigid distortion and local morphological differences. The key challenge
for image registration of section images is to compensate for distortion
between consecutive images that is introduced by the sectioning pro-
cess. Tissue sections are extremely thin (3 to 5 mm) and delicate as a
result. The preparation process (i.e., sectioning, staining, and coverglass
application) can introduce a variety of nonrigid deformations including
bending, shearing, stretching, and tearing. At micron resolutions, even
minor deformations become conspicuous and may prove problematic
when accuracy is critical to the end application. In order to compensate
for such deformations, a nonrigid registration is essential and success
depends on establishing a large number of precise feature correspondences
throughout the extent of the image. This precision requires comparison of
intensity information and is very time consuming with popular comparison
measures such as mutual information (MI).
4. Preservation of 3D morphology of microscopic structures. A specific issue
for 3D reconstruction is the preservation of the 3D morphology of impor-
tant microanatomical structures. Most image registration algorithms focus
on ''optimally'' aligning two images. In principle, they do not take the
integrity of the 3D structure into account. Therefore, the 3D reconstruc-
tion process has to go beyond traditional pairwise image registration by
integrating 3D structure integrity constraint into its overall consideration.

This chapter discusses research efforts in both algorithm design and high-
performance computing (HPC) implementation to address the challenges described
above. A two-stage image registration pipeline designed for aligning large histo-
logical images is presented. The first stage is a fast rigid registration algorithm
based on the matching of high-level features. This approach circumvents the issue
of the presence of numerous and ambiguous local features, providing effective
initialization. In the second stage, nonrigid registration is achieved by precisely
matching a large number of local intensity features using cross-correlation. To