Scalable Image Registration and 3D Reconstruction at Microscopic Resolution - High-Throughput Image Reconstruction and Analysis - page 186

matching can accommodate the full range of misalignments. Third, the feature
descriptions are scalars and are fast and simple to compare once computed. Fi-
nally, many choices of high level features remain mutually distinguishable when
the images to be registered have distinct stain types. This permits, for example, the
alignment of an hematoxylin and eosin stained image with an immunohistochem-
ically stained image.
In contrast, performing corner detection on a typical microscopy image gener-
ates an overwhelming number of features due to the textural quality of the content.
These features are relatively ambiguous, and their comparison requires the use of
neighborhood intensity information and has to account for differences in orienta-
tion and also appearance if distinct stains are used.

High-Level Feature Extraction
Extraction of high-level features is a simple process as the features often corre-
spond to large contiguous regions of pixels with a common color characteristic.
Color segmentation followed by morphological operations for cleanup usually suf-
fice. The computational cost of these operations can be significantly reduced by
performing the extraction on downsampled versions of the original images with-
out compromising the quality of the final nonrigid result. The rigid estimate only
serves as an initialization for the nonrigid stage and a matter of even tens of pixels
difference is insignificant to the outcome of nonrigid stage. Figure 8.1 demonstrates
the extraction process, showing an example from one of our placenta test images.
Descriptions for these features such as size and shape can be computed using tools
like regionprops available in MATLAB's Image Processing Toolbox.

High-Level Feature Matching
Determining an estimate for rigid registration from a set of matches requires a
method that is robust to mismatches. This is especially true in microscope images
where many of the features are guaranteed to have a similar appearance. Regard-
less of how strict criteria for matching features is, it is inevitable that a substantial
amount of coincidental mismatches will occur. The fundamental idea in the ap-
proach presented here is that, given feature sets from the base image { }

b i and

Figure 8.1 Rigid registration for initialization using high level features. (a, b) 5x decimated placenta
image (originally 20x), approximately 4,000

× 4,000 pixels. (c) Regions corresponding to blood
vessels, extracted by color segmentation and morphological cleanup. Correspondences between
elements in the feature sets of the base and float images are used to produce an estimate of the
orientation q and translation T between them. (d) Close-up of (c).