Parallel Feature Extraction - High-Throughput Image Reconstruction and Analysis - page 144

imaging, connections between different functional areas of the brain can be de-
termined. Blow discusses [5] further issues that arise in the creation of a detailed
connectivity diagram of neural tissue. As technology advances, it is possible that
high-resolution maps of an entire brain could be obtained.
Regardless of the imaging method and resolution used, the task of 3D recon-
struction is faced with significant computational challenges, which we term the
computational bottleneck . The computational tasks that need to be performed in-
clude filtering the images to reduce noise, extract relevant features such as contours,
track contours over multiple slices, create a unified representation, and visualize
the structure.
In order to provide a concrete illustration, we examine images gathered from
a technique called serial block-face scanning, introduced by Denk [10].

6.2.1 Serial Block-Face Scanning
In this method, pioneered by Denk and Horstmann [10], the neural tissue is
mounted in a rigid substrate, called a block. The top surface of this block is directly
imaged with an SEM, giving rise to a 2D image. A diamond knife is used to cut
away a thin section (approximately 60 nanometers), which is discarded. The newly
exposed top surface of the block is imaged to give another 2D image at a different
depth. The advantage of this technique is that the successive 2D images are nearly
perfectly co-registered. This overcomes a major difficulty with prior approaches
that required each section to be collected, mounted, and imaged, thus introducing
registration and distortion errors.
The use of this technique gives rise to a 3D image stack , as shown in Figure 6.1.
Once such an image stack is assembled, the next task is to isolate neural structures,

Figure 6.1 A 3D image stack created from the method of Denk and Horstmann [10]. Each 2D
grayscale image is of size 188 × 2,048 pixels. The dataset available online in [10] contains 2,000

such slices. This gives rise to almost 8 GB of image data that needs to be processed.