(247f) Automated Image Analysis of Intracellular Lipid Droplets: Toward High-Throughput Screening of Anti-Obesity Agents

Automated Image Analysis of Intracellular Lipid Droplets:
Toward High-Throughput Screening of Anti-Obesity Agents

Introduction
and Motivation - Cellular
hypertrophy of body fat, or white adipose tissue (AT), underlies many of the
proposed mechanisms for obesity-related illnesses. In hypertrophic adipocytes, the
lipid volume accounts for >90% of cell volume. Thus, an increase in adipose
cellular lipid is equivalent to an increase in cellular volume. This direct
coupling between cellular morphology and metabolism can be exploited to screen
for the effects of agents designed to target specific metabolic pathways by
monitoring the dynamics of cellular lipid droplets. Ideally, the lipid droplets
are monitored noninvasively to enable repeated observations over time, which
can reveal important information regarding the dynamics of lipid droplet
formation.  An attractive approach for monitoring cellular lipid droplets in
cultured adipocytes is to combine optical microscopy and image analysis. In
principle, contrast agents such as lipophilic dyes or labeled antibodies for
lipid droplet-specific proteins are not required, as there is sufficient
contrast between a lipid droplet and the cytoplasm surrounding the droplet. Importantly,
forgoing the use of dyes or labels would not only lessen the experimental
effort, thus aiding high-throughput analysis, but also reduce the likelihood of
introducing additional measurement artifacts into the analysis.

Manual
analysis of microscopy images is both time consuming, error prone, and subject
to biases. This motivates the development of an automated image analysis
algorithm. The major challenges arise from the complexity of the images. In a typical
adipocyte culture derived by inducing preadipocytes or other precursor cells,
differentiation occurs unevenly, leading to a highly heterogeneous population
of cells and lipid droplets.  Due to this complexity, relying on a fixed contrast
thresholds to identify lipid droplets can lead to inaccurate results (Fig 1B).  While
manual tuning of the threshold on an image-by-image basis can improve the
detection of lipid droplets, the improvement is only marginal, and is obtained
at the cost of throughput and potential for bias. In this abstract, we present an
image analysis algorithm that is capable of processing a wide range of images
with little to no user input, and thus well-suited to high-throughput analysis
of lipid droplets in cultured cells.

Image Analysis
Algorithm ? Our algorithm
identifies lipid droplets using four different characteristics of typical
adipocyte images: droplets are (1) relatively light, (2) circular, and (3) surrounded
by a relatively dark boundary, (4) which is also circular. To take advantage of
the first two characteristics, the original grayscale image is converted into
black and white images twice: once using a low threshold, and once using a high
threshold. Instead of choosing a fixed threshold, our algorithm uses a dynamic
threshold value, which adapts to the overall brightness of the entire image. When
the high threshold is applied, the light interior of the lipid droplets and
some light background noise are converted to white, while the rest of the image
becomes black (Fig 2A). When the low threshold is applied, the dark rings
surrounding the lipid droplets and some dark noise are converted to black, while
the rest of the image becomes white (Fig 2B). By inverting the low-threshold
image and filling in the holes, an image results that is white where the dark droplet
boundaries (rings) and dark background noise originally were, and black
everywhere else. An intersection of the two images yields an image that is
white only where the original image was both light and surrounded by a dark
boundary (Fig 2C). In addition to accurately identifying the lipid droplets, these
processing steps also reduce noise in the image, because dark and light noise
cannot occupy the same location in an image. A circle finding algorithm is applied to the low threshold image
(Fig 2D and 2E), and the resulting image is then combined with the intersected
image to generate the final processed image (Fig 1C).

Image Analysis Results The image analysis processing method was
validated against experimental data. We differentiated 3T3-L1 preadipocytes in
12-well plates for 16 days. Every 4 days, we captured 300 images per well, in 6
different wells, then lysed the cells for enzymatic analysis. As expected,
accumulated triglyceride increased over time (Fig 3).  The image analysis and
enzymatic assay data correlated significantly across time point groups (Fig 3A)
and individual wells (Fig 3B).  However, the correlation was stronger for time
point groups (R²=0.99). The weaker correlation for individual wells
is likely due to the heterogeneity of adipocyte differentiation. As the images
captured represent only a fraction of the total culture surface of a well, it
is possible that the captured images do not fully represent all of the cells in
the well.

Conclusions We developed a new image processing method
to support noninvasive analysis of lipid droplets in cultured adipocytes. This
method does not require destruction of the cells for endpoint analysis of cell
metabolism or introduction of exogenous dyes or labels. The image analysis results
correlate strongly with enzymatic assay data, especially for condition group
averages. Prospectively, automated image analysis processing could be use to monitor
the dynamics of lipid droplet formation in cells, and establish high-throughput
screens for agents designed to affect these dynamics.

Figure 1:  Comparison
of currently available image analysis methods to our new algorithm. Original
image of adipocytes in 3D culture (A), processed image using the currently
available program (B), processed image using the new algorithm (C).

Figure
2:  Image analysis algorithm. High threshold image (A), low threshold image
(B), intersected image (C). Circle finding on the low threshold image (D).

Figure 3:
Comparison of triglyceride accumulation assessed via enzymatic assay (x axis)
and image analysis (y axis). The four points represent averages of 6 wells for
day 4, day 8, day 12, and day 16. (A) Each point represents data from a single
well. (B)