![](bg2.png)
Figure 2: A screenshot of KESM Brain Atlas (KESMBA)
(Chung et al., 2011).
However, the tile generation requires time-
consuming manual calibration and the time required
to download a single visualization is significant (~45
to 55 seconds/page for 20 overlays), requiring a lot
of patience on the part of the user.
To address these two issues, we present an
automated image processing pipeline for KESM
mouse vascular images and a parallel multi-scale tile
generation system for web-based pseudo-3D
rendering that includes pre-overlaid tiles. The
system, built on the OpenLayers API, allows full
navigation and multi-scale viewing of the whole
mouse brain data set at maximum resolution using a
conventional web browser.
2 ENHANCED IMAGE
PROCESSING PIPELINE
KESM employs physical sectioning imaging where
thin slices of tissue are concurrently cut and imaged
(Mayerich, Abbott, and McCormick, 2008). These
slices are then re-assembled in order to produce the
final volumetric data set (Kwon, Mayerich, Choe,
and McCormick, 2008). In this section, we describe
an enhanced image processing pipeline that
performs the following tasks:
The Tissue Area Detector detects the portion
of the raw image that contains actual tissue
data.
The Tissue Area Offset Corrector identifies
and corrects errors in the detected tissue area.
The Cropper crops an image based on the
corrected area information.
The Relighter removes lighting artifacts and
normalizes the inter-image intensity level.
The Merger merges multi-column stacks into
a large, single column image
The Overlay Composer generates pre-
overlaid images with a given number of
images (e.g., an overlay of twenty 1μm-thick
images will give a visualization a 20μm-thick
slab) stack.
The Tiler generates tile images for the web-
based map service
In this paper, we provide details for the Tissue
Area Offset Corrector and the Overlay Composer.
The other phases of the pipeline have been described
previously (Kwon, Mayerich, and Choe, 2011).
Automating the image processing steps is critical
for generating brain atlases since the number of
images is extremely large (e.g. 32,792 images in a
whole mouse brain KESM data set). Previously, we
automated key image processing steps including
noise removal, image intensity normalization, and
tissue area cropping (Kwon et al., 2008) (Kwon et
al., 2011). However, the automation of several
important steps remains, including correction of
tissue area detection results. In addition, we
demonstrate that pre-overlaying of images in the
image stack is necessary to improve page load
performance, and must also be automated.
2.1 Tissue Area Offset Corrector
The image processing pipeline starts from the Tissue
Area Detector. A raw KESM image includes blank
regions flanking the region that contains actual
tissue data. Due to the physical sectioning process,
the precise position of the tissue region in each
image can show some variation due to repositioning
of the knife or the objective during extended
cutting/imaging sessions. We previously describe an
automatic method for detecting the tissue region
based on the right-most edge of the tissue (Kwon et
al., 2011). However some images do not have a clear
boundary due to uneven lighting across the knife
edge. Failure to find a proper tissue boundary leads
to incorrect cropping of the images, which are
difficult to manually correct. Such errors impede
proper reconstruction of 3D geometry in subsequent
stages. However, we find that errors can be detected
by observing the computed tissue region in adjacent
images of the image stack. The sum of the difference
between tissue area offsets in neighboring images is
calculated. A sudden spike indicates an improperly
detected tissue area offset. The summation continues
until it reaches a certain threshold C:
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
784