Parallel Processing Strategies for Cell Motility and Shape Analysis - High-Throughput Image Reconstruction and Analysis - page 43

Figure 3.1 General framework for cell shape and behavior analysis.

three 1D convolutions for derivatives and three 1D convolutions for averages in
the corresponding spatiotemporal cubes.
A brute-force implementation where spatial and temporal filters are applied
for each pixel separately within a spatiotemporal neighborhood would be compu-
tationally very expensive since it would have to recalculate the convolutions for
neighboring pixels. For an efficient implementation, the spatial ( x and y ) convo-
lutions are separated from the temporal convolutions, and the 1D convolutions
are applied to the whole frames one at a time. This minimizes the redundancy of
computations and allows reuse of intermediate results.
The spatial convolutions required to calculate I xt , I yt ,

and
I ss where s represents the smoothing filter. Each frame of the input sequence is
first convolved with two 1D filters, either a derivative filter in one direction and
a smoothing filter in the other direction, or a smoothing filter in both directions.
These intermediate results are stored as frames to be used in temporal convolu-
tions, and pointers to these frames are stored in a first-in first-out (FIFO) buffer of
size n FIFO =

and I tt are I xs , I sy ,

n Dt + n At

−

1where n Dt is the length of the temporal derivative filter

Figure 3.2 Cell shape and motility analysis approaches for different datasets.