To evaluate the performance of projection based methods, we acquired test image data by culturing and imaging bone marrow macrophages (BMM). The macrophages isolated from BL6 were cultured on glass cover slip in RPMI medium, supplemented with 10% fetal bovine serum, 100 u/ml penicillin, 100 ug/ml streptomycin, 2 mM GlutaMAX and 50 ng/ml m-CSF (37 C, 5% CO2). The cells were stimulated with LPS 100 ng/ml for 1, 2, 4, 6, 18, and 24 hours, fixed with 3% Paraformaldehyde for 20 min and stained with BODIPY 493/503 (Invitrogen) for lipid bodies, and Sytox (Invitrogen) for nuclei. Unstimulated macrophages as well as the stimulated cells of different time points were imaged with Leica DMIRB confocal laser scanning microscope.
The image stacks form eight groups with varying cell morphologies: two image sets of unstimulated macrophage cells, and a time series experiment with six groups of macrophage images from different time points during the stimulation. For each group, there are five image stacks, each consisting of three channels: 1. fluorescent nuclei 2. fluorescence subcellular stain for lipid bodies also visualizing the cytoplasm and 3. bright field channel. Each of the stacks for every channel consist of 20 individual
-slices. One stack for each channel of the time point
had to be removed because it was erroneously imaged as a single slice instead of a stack. In total, the test data set includes nearly 800 cells.
To enable whole cell segmentation from bright field images, the contrast must be enhanced by increasing the intensity differences between cell and background areas. We achieve this by calculating different measures of variation in the
-direction, projecting the bright field stacks into two dimensional (2-D) images. That is, each pixel in the resulting 2-D projections corresponds to a measure of intensity variation in the
-direction in the original stack in that specific
pixel location. Since there is typically less
intensity variation in the background than in cells, these two classes of pixels can be separated. Specifically, we make the projections using standard deviation (STD), interquartile range (IQR), coefficient of variation (CV), and median absolute deviation (MAD) measures.
The STD projection image is constructed by calculating the standard deviation of intensities in the
-direction for each pixel of the original stack:
is the pixel intensity of
is the mean of the pixel intensities, and
is the total number of
For a more robust measure of variation we calculated IQR projection, the difference between the 75th and the 25th percentiles of the sample. That is, the lowest 25% and highest 25% of the values are first discarded, and the IQR is the range between the maximum and minimum of all the remaining intensities of
In CV projection, the standard deviation of the
-values is divided by the mean of the values
And finally, MAD measures how much “on average” one value deviates from the median of all the values, that is, the median deviation from the median of the intensities of all the
-slices for every
To assess the projections' sensitivity to the number of
-slices imaged for each stack, we applied the STD projection to two different types of reduced stacks, consisting only of three slices. First, the three slices were selected by hand representing nearly the whole
-range of the original stack (slices 2, 10 and 19), referred to as the 3Slices-method. And second, we created five reduced versions of the original stacks by selecting the three slices randomly, referred to as 3SlicesRandom1 to 3SlicesRandom5.
The automated image analysis and cell segmentation for the evaluation of the various projection methods was carried out by the open source CellProfiler software package 
, originally designed for fluorescence microscopy. First, markers for each cell were obtained by detecting fluorescent nuclei with IdentifyPrimAutomatic analysis module. Second, to smooth out small unwanted details from the projections, a Gaussian lowpass filter radius of
pixels was applied by SmoothOrEnhance module. Third, we used the propagation algorithm 
in the IdentifySecondaryAutomatic module for detecting the whole cell areas. For ground truth, the whole cell areas were segmented with the same procedure (excluding the lowpass filter) using fluorescent cytoplasm images to be compared against cell area detection using the various 2-D projections. To simulate a situation where no fluorescent staining is available, the cytoplasmic areas were estimated by an annulus of radius 30 pixels around each nuclei as described, for example, in 
. This estimation approach is referred to as the Annulus-method.
For further validation, we also enumerated fluorescent spots visible in the second fluorescent channel of the stacks. The spot enumeration was done with a kernel density estimation based algorithm 
using a Gaussian kernel. Since this spot enumeration module is not included in the standard CellProfiler distribution, we implemented the analysis pipeline in the Developer's Version of CellProfiler, running on Matlab 2008a. The various approaches for whole cell segmentation are summarized in .
Summary of different whole cell segmentation methods and abbreviations.
We did not discard cells touching image borders, although it is a procedure commonly performed to minimize bias in measurements caused by cells that are only partly visible. These cells allows us to compare segmentation accuracy also on image borders where image quality is often compromised due to nonuniform background. The computational complexity of the analysis is relatively low, taking around 4 seconds per method to calculate the projection and segment the image on a 2GHz PC with Windows Vista.