A Comparison of Methods for Quantifying Silicone Droplets in

A Comparison of Methods for Quantifying Silicone Droplets in Biologics
Using Dynamic Imaging Particle Analysis
FLUID IMAGING
TECHNOLOGIES
®
Lew Brown, William Bernt - Fluid Imaging Technologies, Inc. www.fluidimaging.com
Abstract
Methods
Dynamic Imaging Particle Analysis (DIPA) is increasingly being used as a method of
characterizing sub-visible particulates in protein-based therapeutics. Many of these
formulations, particularly pre-filled syringes, will contain silicone droplets which are
used for lubrication. When reporting particulate content, especially in the 2µm to
10µm size range, it is desirable to eliminate the contribution of the silicone droplets to
the overall particle count.
A Polyclonal IgG sample was obtained (Courtesy University of Colorado) which had
been stressed via agitation. This sample was first run in a FlowCAM® DIPA instrument
unmodified to get a baseline reading for particle load and to verify repeatability of the
data sets, as shown in Figure 3. The sample was then silicone infused by pulling it into
and ejecting it out of a siliconized syringe ten times prior to being run through the
FlowCAM DIPA instrument again. Figure 4 shows the particle size distributions for
five runs of the sample after being silicone infused, again showing good repeatability.
Note also the large increase in overall particle load, caused by adding the silicone and
the additional stressing due to the multiple trips through the syringe prior to the runs.
Since DIPA can measure particle shape parameters, it can potentially be used to “filter
out” the silicone droplets prior to reporting particle count. This poster will look at
several different mathematical approaches to identifying and quantifying the silicone
droplets in the sample. A data set will be acquired with images from a sample protein
therapeutic containing both protein aggregates and silicone droplets. The data set will
first be analyzed visually to make a baseline determination of the number of protein
aggregates and silicone droplets. This baseline data set will have an assigned “particle
type” value given to every particle image collected. This same original data set will
then be subjected to automated, algorithmic analysis using different methods to try to
quantify the silicone droplet content. For each different algorithm or method applied,
the total number of silicone droplets identified by the method will be recorded, and
the number of false positives and false negatives will also be recorded by comparison
to the baseline manually classified data set. These results will allow for a quantitative
evaluation and comparison of each method’s efficacy.
Introduction
In typical particle analysis studies, it is usually desired that the results of automated
analyses be verified for accuracy against a “known” characterization of the material
being analyzed. This almost always means that an analysis/characterization of the
sample be performed by humans using manual microscopy. The human eye/brain
system has the capability to see very subtle differences in images, and to interpret/
interpolate the visual information beyond what a computer system and algorithm are
capable of accomplishing5. This is especially true as the particles get small enough so
that only a few camera pixels are available to form the image itself.
Due to the eye/brain’s ability to discriminate particle
morphology in images beyond where a computer can
successfully accomplish it, it follows that if a human
can not differentiate a particle type from an image,
then certainly a computer can not either. This is a basic
premise of all the work contained herein.
Additionally, while in the past we have seen
publications reporting particle characterization
for DIPA instruments below 2um in size, it is our
contention that any characterization based on particle
shape must be only for particles 2um and larger in size
Note that this limit is for the magnification used in this
work (10X objective), and is based on depth of focus
limitations caused by the fluidics system as opposed
to optical concerns. Figure 1 shows images of silicone
droplets at different sizes, and Figure 2 shows images
in the 1um range that can not be characterized either
by human or especially by computer. More discussion
of this can be found in Reference 5. Note additionally
that this does not say that particles can not be counted
in the 1-2um size range, only that they can not be
characterized by shape reliably in this range. This is
supported by the fact that research papers published on
this same topic also only report particles >2um2,3,4.
Figure 1: Silicone droplet particle images
zoomed 8x with binary edge overlaid in
bottom image.
Figure 2: Nondescript particle images
zoomed 8x with binary edge overlaid in
bottom image.
The major difference here over a value filter is that the statistical filter introduces
“weighting” of the particle properties based upon the statistics, so that the properties
that are most influential in making the training set unique are weighed heaviest in the
resulting analysis. So a statistical filter was built from one of the data sets by selecting
several different silicone droplet particle images as a training set, and applying it to
each of the runs as a filter. An increase of 22% more silicone drops was found over the
Circularity (Hu) filter above using this technique.
Stressed Polyclonal IgG (before silicone)
100,000
10,000
Run 1
1,000
Run 2
Run 3
100
Run 4
Run 5
10
Detection and enumeration of silicone droplets in biologics formulations (prefilled syringes, etc.) is important because the identified silicone droplets should be
eliminated from overall particle counts which may be reported under USP <788> or
other internal/external requirements prior to reporting. Different methodologies for
numerically separating the silicone from the protein aggregates have been reported in
the past1,2,3,4, with varying degrees of success.
evaluated against that value range and either are within it or not, therefore a member
of the class or not. A more sophisticated way of doing filtering is know as “statistical
filtering”. In statistical filtering, several particles are selected (called a “training set”) by
the user which represent the “type” the user wishes to find (based on eye/brain input).
The statistics of all measurements for the selected particles are then compared against
the statistics for each particle in the run to determine how similar they are to the
selected particles. Typically, this can be done by examining how close in n-dimensional
space (where n = the number of particle properties measured) the target particle under
study is to the original selected ones7.
Results
1
2‐5um
5‐10um
10‐25um
25‐50um
50‐100um
100+um
Figure 3: Five replicate runs of original Polyclonal IgG prior to being infused with silicone.
Stressed Polyclonal IgG with Silicone Oil
100,000
10,000
1,000
Run 1
Run 2
Run 3
100
Figure 5: Screen capture of the hand-classifying process. Particle #10 in the right-hand image window has been
selected and is being added to the class “Protein” in the Training Set window (lower left), thereby appending a
parameter “Type (known) = Protein” to the data associated with Particle #10.
Run 4
Run 5
For the hand classification, four particle types were created: “protein” aggregate,
“silicone” droplet, “combination” and “unknown”. The unknown had to be created
because, even at 2um and larger, there will still be some particles that just cannot
be classified, even by a human (Figure 6 shows some examples). The “combination”
category consists of both multiple silicone droplets stuck together (forming a
different shape than singlets) and
predominantly silicon droplet(s)
embedded into protein aggregates
(Figure 7 shows some examples). This
is not an uncommon occurrence4,
and was probably exacerbated by the
multiple runs through the syringe (an
area suggested for future modification
Figure 6: Typical examples of Particles hand-classified as
in this continuing research).
“Unknown”. Original size on left, zoomed 8X on right.
The most commonly used filter
10
criterion for separating silicone
droplets from protein aggregated
is aspect ratio (width/length). This
1
2‐5um
5‐10um
10‐25um
25‐50um
50‐100um
100+um
is because the silicone droplets are
usually perfectly spherical in shape
Figure 4: Five replicate runs of Polyclonal IgG after silicone infusion.
versus the amorphous shape quality
of the protein aggregates. However,
Figure 7: Typical examples of Particles hand-classified as
“Combination”. Original size on left, zoomed 8X on right.
At this point, an enhanced development version of the FlowCAM VisualSpreadsheet® aspect ratio is a relatively crude
software was used to “hand-classify” each of four runs of the protein/silicone mix. The measurement, and, even though the
procedure used was to view every particle image in the data set, and have the operator general shape is amorphous, it is easy to find protein aggregates having a high aspect
then append a particle “Type (known)” to each particle’s spreadsheet information as an ratio, producing false negatives. Several different cut-off values for aspect ratio filters
additional data field. Figure 5 shows a screen capture of this procedure. For the reasons were tried, and it was found that Aspect Ratio ≥0.85 produces the best results.
noted above, this was only done for the data set >2um in size, which cut the number of
particles to be appended by about half since the largest size bin is the 1-2um one. Even Various different measurements of particle circularity are available in the
with this reduction in data, it left an average of 2,500 particles to be hand-classified,
VisualSpreadsheet software, but one that is particularly effective for detecting silicone
which takes considerable time.
droplets is “Circularity (Hu)”. Hu Circularity has been shown to produce better results
for circularity than other measures when looking at particles with boundary defects,
Once the hand-classification is finished, we now have each data set containing an
i.e. when the perimeter of the circle is not perfect6. This is exactly what occurs when
additional field for each particle of “Type (known)”, which can be sorted on by
silicon droplets are imaged with very few pixels, therefore at the limits of the optical/
VisualSpreadsheet. At this point, any time we use a filter to attempt characterizing the sensor system (see Figure 1). Once again, different thresholds for Circularity (Hu) were
number of silicone droplets, we can then sort the filter results by “Type (known)”, and tried, with a value Circularity (Hu) >0.95 producing the best results, an increase of
thereby determine the number of particles correctly identified as silicone, the number 42% more silicone drops found over the best results using an aspect ratio filter.
of false positives and the number of false negatives. The result give a quantification of
the efficacy of the filter in terms of how it did against the hand-classified results.
The previous filters discussed are what is called “value filters”, meaning that a range
of values for a certain particle measurement are defined, and then all particles are
The last filter tried was a customized filter derived from the image statistics. To do
this, we first made a data set consisting of only the silicone droplets for each run which
had been hand-classified with “Type (known)” of silicone. We then took the summary
statistics for this data set, and sorted the results based upon the Coefficient of
Variability (CV) for each of the parameters, from lowest to highest. The thought is to
take the image parameters from the silicon with the lowest CV, and use their min/max
to construct a customized value filter. We ended up using a combination of particle size
(ESD), Circle Fit and Circularity (Hu). On running this filter, we got our best results
yet with an increase of 34% more silicone drops found over the statistical filter results
above.
Figure 8 shows a summary of the results found for each filter in terms of the increase
in number of silicone droplets correctly identified by the filter, using the Aspect Ratio
>0.85 as a baseline.
Filter: Aspect Ratio ≥ 0.85
Filter: Circularity (Hu) ≥ 0.95
Filter: Statistical Filter
Filter: Customized Value Filter
% Increase Overall vs. Baseline (Aspect Ratio filter)
N/A (baseline)
42%
73%
130%
Figure 8: Summary of efficacy results for different silicone droplet filters, using the Aspect Ratio filter as the baseline,
showing the total increase over the baseline.
Conclusions
The ability to actually see each particle image in order to hand-classify them is critical
in order to be able to truly understand the efficacy of software filters to segment out
an individual population of particles in a sample, in this case silicone droplets. The
method used in this poster shows great promise in yielding real insight into how well
a particular method may be at quantifying silicone droplets in protein-based therapies.
The results presented here represent only a beginning at using this methodology to
better understand the use of DIPA in characterizing biologics.
Having a larger variety of particle properties from which to choose from when
constructing filters increases the chance of finding those properties that can best
identify a particular type of particle. It is clear that more advanced measurements,
such as Hu Circularity, and more advanced methods, such as statistical filtering,
can significantly increase the accuracy of silicone quantification in protein-based
therapeutics.
References
1. Sharma, D.K., Oma, P., & Krishnan, S, (2009). Silicone Microdroplets in Protein Formulations-Detection and
Enumeration. Pharmaceutical Technology, 33 (4), 74-79.
2. Strehl, R., Rombach-Riegraf, V., Diez, M., Egodage, K., Bluemel, M., & Koulov, A.V. (2011). Discrimination
Between Silicone Oil Droplets and Protein Aggregates in Biopharmaceuticals: A Novel Multiparametric Image Filter
for Sub-visible Particles in Microflow Imaging Analysis. Pharm Res, 29 (2), 594-602.
3. Weinbuch, D., ZÖlls, S., Wiggenhorn, M., Friess, W., Winter, G., & Hawe, A. (2013). Micro-flow imaging and
resonant mass measurement (Archimedes)--complementary methods to quantitatively differentiate protein particles
and silicone oil droplets. J Pharm Sci, 102 (7), 2152-2165.
4. Gerhardt, A., Mcgraw, N.R., Schwartz, D.K., Bee, J.S., Carpenter, J.F. & Randolph, T.W. (2014). Protein
Aggregation and Particle Formation in Prefilled Glass Syringes. J Pharm Sci, 103 (6), 1601-1612.
5. Brown, L. (2014). Can a Computer Characterize a Particle My Eye Can Not? URL http://www.particleimaging.com/
can-a-computer-characterize-a-particle-my-eye-can-not/ .
6. Žunic, J., Hirota, J., & Rosin, P.L. (2008). A Hu moment invariant as a shape circularity measure. Pattern
Recognition, 43 (1), 47-57.
7. Brown, L. (2008). Particle Image Understanding – A Primer. URL http://www.fluidimaging.com/resource-centerwhitepapers.htm .