Mouse Embryo Visual Project   

INTRODUCTION

To understand the mechanisms of human development, embryologists study both normal development and development that leads to the birth defects. One way to gain further understanding of the development of an embryo is through 3-D reconstruction from serial sections taken from an optical microscope. The Mouse Embryo Visual Project (MEVP) is intended to develop a semi-automatic system utilizing the concepts of computer graphics and 3D visualization and at the same time allow to explore the anatomy of a mouse embryo. The modern segmentation method based on the modification of 'snakes' will be used to speed up the process.

CONTOUR-BASED 3D RECONSTRUCTION

The contour-based reconstruction, begins with tracking of contours that outline structures at a series of slices. The traced contours are stored in a database, and various methods can be use to render the contoured object. This reconstruction method is useful for observing the shape of the object of data compression because the 3-D volume of a structure is represented simply by a surface created by forming tiles between the adjacent contour lines. Contour based 3-D reconstruction is widely used with light microscope sections, and with conventional thin electron microscope sections. The contour-based reconstruction is also the method of choice for reconstructions where there are significant discontinuities between structures on adjacent sections. It is also useful for morphological measurements. Both commercial and otherwise available software for 3D reconstruction of MRI, CT, confocal and serial-secion data for medical sciences is listed at http://biocomp.arc.nasa.gov/3dreconstruction/software/. See 3D Reconstruction homepage for more information.

The alternative to contour-based reconstruction is volume rendering. There are several software packages available for volume rendering. Images obtained from volume rendering contain the information inside the object, and they are useful for over viewing the object. However, it is difficult to observe the precise shape of the object from images, due to the difficulty of determination which point in volume is actually contributing to the final projection.

THE TECHNIQUE FOR ACQUIRING THE MOUSE EMBRYO DATA

We are going to observe a mouse embryo including its inner structures at development stage approximately 10.5 days old.
mouse embryo before processing
Photo of a mouse embryo before processing.
The size of the embryo in so early development stage is very small, i.e., about 4.5 mm in height. A normal way to observe the inner structures of a microscopically small spacemen is to use cutting sections from the spacemen. The mouse embryo, in our case, was further sectioned with an ultramicrotome at an average thickness of section about 7 micrometers. This gives us 636 cross-sections that are further processed.

If alignment of the sections is exact, close spacing will show accurate detail in the 3-D reconstruction. However, sections cut from biological spacemens are often slightly distorted, particularly if using ultra-thin sections. Sections cannot then be completely aligned across their full width and this may result in spurious detail and a false impression of the morphology. It may be possible to correct a distortion by manipulation of the images or the data. To solve this problem we explore the image registration techniques over serial cross-sections.

Data Preparation

Making a microscope preparation from mouse embryo is a very complex problem, which consist of following steps: Each cross-section is mounted on glass and then scanned at magnification of 74 times by an optical microscope, to which is attached a photo camera. Images taken from the photo camera are stored in high-resolution compact discs with after conversion into digital images. The resolution of digital images was selected to be 720x580. We have experimented with a CCD video camera connected to graphics work-station and attached to the microscope. The resolution of the CCD video camera is not so high. For comparable resolution with a photo camera, we need to use a HDTV (high vision) video camera.
mouse embryo cross-section
Scanned mouse embryo cross-section. It shows embryo's head and brain.

Image Registration

Registration refers to the alignment of data from the same or different sensors. Alignment of sections from serial microscopy is required to compensate for combination of the two, misregistration and physical changes over the space. Sections from serial microscopy are not only translated and/or rotated. Often they are deformed due to the heating required in the preparation of the tissue for microscope slices. Thus, serial microscopy requires a combination of a linear transformation to bring the sections into approximate alignment, followed by a nonlinear transformation to account for distortions in the tissue preparation process.

To reconstruct the three-dimensional structure of the mouse embryo from the set of cross-sections, all sections must share a common reference coordinate system. One approach for finding the approximation of the reference coordinate system is as follows: The entire mouse embryo is embedded into a paraffin block to make the sectioning process easier. Four marks can be made in the paraffin with four shots of laser and consequently filled with gelatin. The gelatin marks on each section determine the reference coordinate system that can be approximated using the image registration techniques. However, because important data images were already taken without marks, no reference marks were adopted. Therefore, the image registration process is necessary to find a good approximation of the reference coordinate system.

In the majority of reconstructions the images can be satisfactory aligned by eye. It is possible to establish the best fit for the three variables, x- and y- shift and rotation. An alternative of manual technique for aligning two images, is that the user translates and rotates one of the images while the other is held stationary. This method uses color combination instead of motion used in. The two images are colored using two distinct colors, and images are added to form a third color. The best alignment is easy to recognize because it maximizes the area of the third color.
 
  
Static image.
Consecutive image.
Color merging.
image registration
The best registration.
Vectors and landmarks shown in above images determine the translation and rotation of images when two vectors align.

The series of sections is registered by repeating the registration of image pairs from top to bottom or vice versa. We should remember that small error of alignment can accumulate a significant shift or rotation over many sections. It is therefore necessary to register the actual section also to the sections that are far away.

CONTOUR EXTRACTION FROM IMAGE DATA

The large morphometric variability in biomedical organs requires an accurate fitting method for a pregenerated contour model. We are seeking the outline contours of a different shapes visible in the cross-sections. We proposed and implemented a physically based approach to the fitting of two-dimensional contours using texture feature vectors. This method is an extension of well known 'snake' model allowing even automatic contour splitting into several parts. (Images related to the snake model.)

Unfortunately, the necessary numerical iteration process requires a good approximation of an initial step solution. Generally, the approximation is carried out interactively tracking the contour using a mouse locator near the edges within the image cross-sections. This initialization need not be performed for each contour and all sections. Setting the initial position near the edges of a target object only for a few selected cross-sections is sufficient. Results from one section are automatically used in the next cross-section as initial position. Continuing this process for all cross-sections we will finally have a stack of contours without knowing how are they related. This is an example of what is the result in this step:

Stack of extracted contours.

SURFACE FORMATION FROM CONTOUR DATA

First, the correspondence problem must be solved between multiple contours in the adjoined sections. Last, the triangulation between each pair of contours must be performed. The result of the triangulation is a surface enveloping the stack of extracted contours called "geometric" model. Such a geometric model is built for each structure of interest.

The next step is to present the models in a way that allows people to understand the shape and relationship between models. This is done by visualization techniques. Now, you can see several images showing different parts of a mouse embryo.

Reconstructed mouse embryo
Reconstructed mouse embryo.

© Roman Durikovic 1998