Image Registration

If we want to do 3D modeling, we first need to know how to translate 2D sections into 3D objects. For this, we use image registration to align sections of tissue into the same coordinate plane.

Registration vs. Native 3D Microscopy

Microscopy gives us a view of biology on the scale of microns. A typical brightfield microscopy section at 40x optical magnification is digitized at about 0.25 microns per pixel (mpp). For comparison, a cell nucleus is on the order of 10 microns in diameter.

While some types of microscopy are inherently 3D (confocal microscopy, for example), these techniques are often expensive, require specialized preparation, and are limited in their ability to visualize “deep” tissue sections. Further, the majority of clinically-available samples are not prepared with 3D microscopy in mind: they are non-specifically stained (e.g. with hematoxylin and eosin) and intended for use in brightfield settings.

Macro-scale Structure from Micro-scale Data

We can combine the benefits of 2D microscopy (high resolution) with those of 3D microscopy (realistic biological structure) by obtaining multiple serial sections of tissue. This is done by sequentially cutting a block of tissue, fixing each slice on a glass slide, and staining for the structures of interest. These slides are then scanned at a typical working resolution (between 0.25 - 0.5 microns per pixel) to yield a large set of data for each tissue sample.

These images are in sequence, but not yet in alignment. To line up our sequence of images, we must register them so they can be stacked together to reconstruct the underlying 3D architecture. That process of figuring out how to transform image A so that it aligns with image B is the goal of registration methods.

Improving Existing Registration Techniques

There are many methods for performing registration; interested readers are encouraged to check out Fiji, a package of ImageJ that contains popular registration methods. In our lab, we are interested in expanding the ability of these algorithms to tackle the following problems:

1. Large images: Whole Slide Imaging (WSI) leads to tissue images that can measure several gigabytes of space (corresponding to images of dimension 100,000 x 100,000 and larger). Efficient registration of these images would mean that we can create stacks covering a wider area of biological tissue.
2. Deep stacks: Tissue sections can average around 6 microns of thickness. This means that serial sectioning a block of tissue that is 1 centimeter deep may result in up to 10,000 sections! If the registration algorithm cannot handle errors across several images, then those errors may become magnified across such a large space.
3. Missing / Noisy Data: As any lab technician will tell you, getting 10,000 perfect slides is impossible. Tissue tearing, inconsistencies in staining, and lost sections are a reality that must be handled at the registration side. The algorithm must be robust to these changes; also, it must be able to interpolate between sections that are lost.

We aim to build on existing methods to create a robust set of techniques that are driven by the data itself: by understanding the scale of the data and structures of interest, it should be possible to tune the parameters of these methods to yield accurate registered stacks that reflect the underlying volume.

• » Next CT Reconstruction