Parallel Processing Strategies for Cell Motility and Shape Analysis - High-Throughput Image Reconstruction and Analysis - page 39

CHAPTER 3
Parallel Processing Strategies for Cell
Motility and Shape Analysis
3.1 Cell Detection
The first step in cell behavior analysis is cell detection. Cell detection returns an
initial contour close to actual cell boundaries, that is later refined by the cell seg-
mentation process. A wide range of techniques have been applied for detecting
and segmenting biological objects of interest in video microscopy imagery includ-
ing spatially adaptive thresholding, morphological watershed, mean shift, active
contours, graph cuts, clustering, multidimensional classifiers like neural networks,
and genetic algorithms. Various factors such as type of the cells, environmental
conditions, and imaging characteristics such as zoom factor affect the appearance
of cells, and thus choice of detection method. But in general, features used in
cell detection can be grouped into three categories: intensity-based, texture-based,
and spatio-temporal features. Intensity-based detection approaches (i.e., intensity
thresholding [1, 2] or clustering [3]) are suitable for lower resolution or stained
images where cells or nuclei appear semi-homogeneous and have distinct intensity
patterns compared to the background. For high-resolution unstained images where
interior of the cells appear highly heterogeneous or where interior and exterior in-
tensity distributions are similar, features based on spatial texture (i.e., ridgeness
measures in [4]) or features based on spatiotemporal features or motion are needed.
In the following section, we describe the flux tensor framework for accurate
detection of moving objects (which in this context correspond to moving cells)
in time lapse images that effectively handles accurate detection of nonhomoge-
neous cells in the presence of complex biological processes, background noise, and
clutter.
3.1.1 Flux Tensor Framework
The classical spatiotemporal orientation or 3D grayscale structure tensor has
been widely utilized for low-level motion detection, segmentation, and estima-
tion [5], since it does not involve explicit feature tracking or correlation-based
matching. The traditional structure tensor fails to disambiguate between station-
ary and moving features such as edges and junctions without an explicit (and
expensive) eigendecomposition step at every pixel to estimate a dense image veloc-
ity or optical flow field. Flux tensor successfully discriminates between stationary
and moving image structures more efficiently than structure tensors using only the
trace.
39