The use of biomedical imaging in scientific research continues to grow, making it critical to efficiently
utilize the enormous potential contained in these 3D images. Such images are typically generated
either from a stack of tissue cross-sections imaged by 2D methods such as light microscopy, or as
volumetric grids imaged by 3D methods such as MRI or CT. Two such examples are shown below. As these imaging modalities become
more commonplace, scientists are moving from studying single subjects in isolation to comparative
analysis of images obtained from multiple subjects, or from the same subject over time. Our research in this direction
addresses a fundamental issue in multiple subject studies: How to place the information of one subject
in the context of others in order to support cross-subject biological or clinical hypotheses.
To perform subject-variant or time-variant analysis, we propose to organize the image data using a geometric, deformable atlas. The atlas allows fast and accurate registration between images as well as efficient, multi-resolution queries among a large pool of 2D or 3D data.
2D Prototype - GeneAtlas [SGP 2003, PLoS 2005]
As a pilot project, we have developed a 2D atlas-based database for expression images of
a small number of genes in P7 mice and made it available online at www.geneatlas.org.
features a set of web-based tools that support comparison of 2D gene expression images corresponding
to selected sagittal cross-sections of the mouse brain. The key to performing such a comparison
is to construct an atlas, modelled as a quadrilateral
subdivision mesh, for each key sagittal cross-section of the mouse brain. This atlas partitions the brain into
disjoint anatomical regions, and is deformed onto each expression image in a semi-automatic manner
so that the expression data on the image is stored on the anatomy-based coordinate system associated with
the atlas. We have recently improved the deformation technique to adopt a statistical approach to further reduce human intervention [TMI 2007].
The resulting annotated atlas forms a 2D spatial database that allows accurate comparison
of expression data on a common, anatomically-based coordinate system.
Using the convenient online tools that Geneatlas.org provides, biologists can specify spatial queries
into the database and seek solutions to biological problems. In one example shown in the figure,
a user has selected a portion of the midbrain containing the substantia nigra (the highlighted area)
and then asked for those genes whose expression patterns are similar to the expression pattern for the
gene Slc6a3, a gene central to dopamine transport. The right of the figure shows the eight expression patterns
that are most similar to that of Slc6a3 (out of 120 gene currently in the database). While seven of
these genes are known to be related to dopamine transport, the gene Hbn1 has a completely unknown
function. Thus, this search suggests that further investigation into the relationship between Hbn1 and
dopamine transport is in order. There is a potential medical significance to this, since Parkinson's
disease is characterized by a loss of neurons in the substantia nigra.
Creating 3D atlases|
A key step in extending our 2D approach to 3D is to build a 3D atlas that is suitable for
storing and querying spatial information (such as gene expressions). While many representations have been proposed for 2D atlases of
anatomical structures, we propose to use volumetric subdivision mesh
as our 3D atlas, which has a multi-resolution, partitioned structure with smooth interior parameterizations. Such task is especially challenging when the input data to build the atlas of consists of a stack of 2D images that may contain distortions and incoherences. Using the mouse brain as an example, we address this challenge in three steps:
- 3D assembly of 2D stacks of coronally sectioned images [JNM 06] The input to our method is a stack
of high-resolution optical sections of the brain. Sectioning of the brain introduces
small non-linear distortions into the tissue in the direction of the sectioning cut. Using novel image warping
developed an automated method that compensates for sectioning distortions and assembles
these 2D images stacks into a single 3D volume.
- Surface construction from planar contours [PG 2005, EG 2008]
We next developed an automated method that takes 2D curves separating multiple anatomical regions on parallel cross-sections and creates a high-quality 3D surface. In the mouse brain example, we had a mouse anatomist trace the contours of up to 20 anatomical regions on each distortion-corrected tissue section. Our method then creates a triangulated
surface that models the partitioning of the brain into 3D anatomical volumes. The constructed surface
is guaranteed to be geometrically correct (e.g., no self-intersections and gaps), and the topology of the surface
can be easily modified by an anatomist. More recently, we have extended this approach to construct smooth, closed surfaces from contours defined on non-parallel planes.
- Volumetric subdivision atlas construction
Our next step is to apply a polygon reduction method to the resulting surface model and tetrahedral-
ize the interior of these reduced surfaces. The resulting
tetrahedral mesh will be treated as the base mesh for the tetrahedral volumetric subdivision
scheme. Here we show an example of a prototype atlas of a human foot bone, which has been used in analyzing
bones density variations in Diabetes patients [JDI 2008]
Towards a 3D Database|
High-throughput registration of gene expression images
Given the 3D atlas, we next need to develop methods for
registering stacks of 2D gene expression images onto this 3D atlas. While there has been substantial
work on registration, registering gene expression images is
particularly difficult due to the variability of the staining patterns produced by the RNA probes
and the local tissue distortions induced by the brain sectioning process. We shall use methods based on machine learning to
construct filters that normalize gene
expression images and then register a 3D atlas to a stack of these images using methods developed
for fitting subdivision meshes to scattered data generated by laser range finders.
Analysis and exploitation of the spatial database
The atlas technology and image registration
methods can be used to construct a 3D spatial database consisting of a 3D atlas annotated by
the spatial expression patterns for thousands of genes. To aid in the analysis of patterns in
this database, we need to develop convenient web-based interfaces that allow users to compare
and cluster 3D gene expression patterns. As a trial project, we intend to examine the spatial
clustering of expression patterns for genes that encode the Wnt family of proteins. These proteins
constitute signaling cascades and play a pivotal role in cancer and embryonic development.
A Geometric Database For Gene Expressions Over The Mouse Brain
Invited talk at Mallinckrodt Institute of Radiology, Washington University, June 2006.
(Download PPT 59MB)
Building 3D surface networks from 2D curve networks with application to anatomical modeling
Paper presentation at Pacific Graphics, Macau, China, October 2005.
(Download PPT 20MB)
A Spatial Database of Gene Expression Patterns for the Mouse Brain
Invited talk at the Colloquium Series, CIIT Centers for Health Research, Raleigh-Durham, April 2004.
(Download PPT 45MB)
Assembling Image Stacks for Sectioned Tissue into Coherent 3D Images
Invited talk at the Annual Workshop of Deformable Modeling ITR, Rice University, March 2004.
(Download PPT 27MB)
Geometric Database of Gene Expression for the Mouse Brain
Paper presentation at Eurographics Symposium on Geometry Processing, Aachen, Germany, July 2003.
(Download PPT 6MB)
Subdivision Method in Gene Expression Analysis
Invited talk at the Annual Workshop of Deformable Modeling ITR, Stanford University, Feburary 2003.
(Download PPT 866KB)