Name: | Description: | Size: | Format: | |
---|---|---|---|---|
457.34 KB | Adobe PDF |
Advisor(s)
Abstract(s)
Abstract—Three-dimensional datasets from biological tissues
have increased with the evolution of confocal microscopy. Hepatology
researchers have used confocal microscopy for investigating
the microanatomy of bile ducts. Bile ducts are complex
tubular tissues consisting of many juxtaposed microstructures
with distinct characteristics. Since confocal images are difficult
to segment because of the noise introduced during the specimen
preparation, traditional quantitative analyses used in medical
datasets are difficult to perform on confocal microscopy data
and require extensive user intervention. Thus, the visual exploration
and analysis of bile ducts pose a challenge in hepatology
research, requiring different methods. This paper investigates
the application of unsupervised machine learning to extract
relevant structures from confocal microscopy datasets representing
bile ducts. Our approach consists of pre-processing,
clustering, and 3D visualization. For clustering, we explore
the density-based spatial clustering for applications with noise
(DBSCAN) algorithm, using gradient information for guiding
the clustering. We obtained a better visualization of the most
prominent vessels and internal structures.
Description
Aiming at a better result from previous works, we employed
some heuristics found in the literature to determine
the appropriate parameters for the clustering. We proposed
our methodology by adding some steps to be performed
before the clustering phase: one step for pre-processing
the volumetric dataset and another to analyzing candidate
features to guide the clustering. In this latter aspect, we
provide an interesting contribution: we have explored the
gradient magnitude as a feature that allowed to extract relevant
information from the density-based spatial clustering.
Besides the fact that DBSCAN allows easy detection of
noise points, an interesting result for both datasets was
that the first and largest cluster found as significant for the
visualization represents the structure of interest. In the red
channel, this cluster represents the most prominent vessels,
while in the green channel, the peribiliary glands were made
more evident.
Keywords
Confocal microscopy data Image processing DB-scan clustering Volumetric visualization