Image Analysis Made Easy with
ImageStream Cytometry
April 27, 2011
Thaddeus George, Ph.D.
Director of Biology
Amnis Corporation
™
The ImageStream® System
High speed imaging (ImageStream™)
+
Quantitative image analysis (IDEAS®)
=
Advances your research
But WHY???
™
The ImageStream® System
High speed imaging (ImageStream™)
+
Quantitative image analysis (IDEAS®)
=
Advances your research
Because you can discriminate cells
objectively and statistically based on their appearance
™
ImageStream Quantitation: How do I
measure what I see?
Incomplete NF-κB translocation in LPS-stimulated monocytes
™
ImageStreamX
• 1,000 cells per second
• 12 image channels per cell: SSC, brightfield, fluorescent
• Up to 5 lasers (488, 405, 561, 592, 658 nm)
• 430-800 nm imaging bandwidth
• Multiple magnifications (60X/.9NA, 40X/.75NA, 20X/.5NA)
• AutoSampler for 96 well plates
• Extended depth of field optics (EDF)
™
IDEAS® Analysis Software
™
IDEAS® Analysis Software
• High content
morphometric analysis of
tens of thousands of
images
• Application wizards for
validated protocols
• “Automated feature
finder” quickly finds the
best feature for analysis
• 85 parameters per
channel, 16 masks, and
user defined features for
advanced image-based
discrimination of cells
™
ImageStream Quantitation: How do I
measure what I see?
-2.2
+4.0
-1.9
+3.3
-1.5
+3.1
-1.4
+2.8
-0.9
+2.6
-0.5
+0.1
+0.5
+0.9
+1.4
+1.8
™
ImageStream Quantitation: How do I
measure what I see?
Dose response and time course: LPS-stimulated monocytes
™
ImageStream Quantitation: How do I
measure what I see?
Dose response and time course: LPS-stimulated monocytes
™
ImageStream Quantitation: How do I
measure what I see?
Dose response and time course: LPS-stimulated monocytes
™
ImageStream Quantitation: How do I
measure what I see?
75
Percent Translocated
Percent Translocated
Dose response and time course: LPS-stimulated monocytes
50
25
100
75
50
EC50 = 6.31 ng/ml
25
0
0
0
15
30
45
60
Time (min)
75
90
-8
-7
-6
-5
-4
-3
-2
Log [LPS] (mg/ml)
™
How do I best analyze my data?
• IDEAS provides 100s to 1000s of features to choose from
• 85 base features per image (size, shape, texture, signal
strength, location, comparison)
• 16 masking algorithms
• Ability to combine masks/features
• Advantage = powerful ability to discriminate different cell
types based on their imagery
• Challenge = how to pick the best feature(s) that provide the
best statistical separation between cell types
ImageStream Data Interpretation
™
ImageStream as a tool for
discriminating cells
Challenge = how to pick the best feature(s) that provide the
best statistical separation between cell types
Task
Design / execute experiment
High speed imaging
Identify cells (“truth”)
Calculate features and statistics
Rank features by discriminating power
Interpret / refine result
Human ImageStream
X
X
X
X
X
X
™
Finding the best translocation analysis
Challenge = how to pick the feature that best discriminates
untreated and LPS-treated cells
Task
Design / execute experiment
High speed imaging
Identify cells (“truth”)
Calculate features and statistics
Human ImageStream
X
X
Rank features by discriminating power
Interpret / refine result
™
Finding the best translocation analysis
Incubate cells with /
without LPS for 1 hr
Fix and perm cells. Label
NFkB with FITC and stain
nuclei with DRAQ5
Collect large event image
files on the ImageStream
Quantify nuclear
translocation of NFkB
Which feature is best at distinguishing
the samples?
™
Finding the best translocation analysis
Challenge = how to pick the feature that best discriminates
untreated and LPS-treated cells
Task
Design / execute experiment
High speed imaging
Identify cells (“truth”)
Calculate features and statistics
Human ImageStream
X
X
X
Rank features by discriminating power
Interpret / refine result
ImageStream Data Interpretation
™
Truth sets: untreated vs LPS-treated cells
Untreated
LPS-Treated
™
Finding the best translocation analysis
Challenge = how to pick the feature that best discriminates
untreated and LPS-treated cells
Task
Design / execute experiment
High speed imaging
Identify cells (“truth”)
Calculate features and statistics
Rank features by discriminating power
Interpret / refine result
Human ImageStream
X
X
X
X
X
X
™
Finding the best translocation analysis
• Four Features for translocation:
• Brightfield Area
• Measures change in cell size
• Not expected to be useful
• H Variance Mean_NFkB
• Measures change in NFkB texture or pattern
• Does not require mask
• Nuc:Cytoplasmic NFkB Ratio
• Measures comparative amount of NFkB in the nucleus and
cytoplasm
• Most intuitive
• Similarity_NFkB/DRAQ5
• Measures increasing similarity of NFkB and DRAQ5 images
as NFkB moves to the nucleus
• Amnis claims this is the best feature
™
Finding the best translocation analysis
Area_BF
Variance_NFkB
ImageStream Data Interpretation
Nuc:Cyt NFkB
Similarity
™
Finding the best translocation analysis
Challenge = how to pick the best feature(s) that provide the
best statistical separation between cell types
Task
Design / execute experiment
High speed imaging
Identify cells (“truth”)
Calculate features and statistics
Rank features by discriminating power
Interpret / refine result
Human ImageStream
X
X
X
X
X
X
™
Rd - statistical measure of discrimination
Fisher’s Discriminant ratio (Rd)
• Measure of the statistical separation a feature provides
between two populations (1&2) using their means and
standard deviations:
Rd = (Mean1-Mean2) / (StdDev1+StdDev2)
• Note that discriminating power increases with increasing
mean differences and decreasing standard deviations
• Populations 1&2 can be:
• Different samples
• Different hand-picked truth sets
• Same set of cells analyzed with different features
ImageStream Data Interpretation
™
Finding the best translocation analysis
Area_BF
Rd = -0.2
Variance_NFkB
Rd = 1.6
Nuc:Cyt NFkB
Similarity
Rd = 2.4
Rd = 3.3
• Similarity thus provides the best statistical discrimination
between the untreated and treated groups
• Results can be refined with protocol development
™
Finding the best translocation analysis
Conclusion: Similarity provides the best statistical
discrimination between the untreated and LPS-treated groups
Task
Design / execute experiment
High speed imaging
Identify cells (“truth”)
Calculate features and statistics
Rank features by discriminating power
Interpret / refine result
Human ImageStream
X
X
X
X
X
X
™
Classification of cells based on
actin distribution
Incubate cells with
polarizing stimulus
Fix and perm cells. Label
actin with FITC and stain
nuclei with DAPI
Collect 20,000 event
image files on the
ImageStream
Classify: round vs
elongated cells with
polarized vs uniform actin
Which actin feature(s) are best at
classifying the cells?
™
Classification of cells based on
actin distribution
Challenge = how to pick the feature(s) that best discriminate
cells with differences in actin distribution
Task
Design / execute experiment
High speed imaging
Identify cells (“truth”)
Calculate features and statistics
Human ImageStream
X
X
X
Rank features by discriminating power
Interpret / refine result
ImageStream Data Interpretation
™
Define cell types based on actin
distribution
™
Define cell types based on actin
distribution
• 2 classifications
• Polarized vs Uniform actin (actin polarization)
• Round vs Elongated cells (cell shape)
™
Truth set #1: Uniform vs Polarized actin
Uniform actin
Polarized actin
™
Classification of cells based on
actin distribution
Challenge = how to pick the feature(s) that best discriminate
cells with differences in actin distribution
Task
Design / execute experiment
High speed imaging
Select truth sets
Calculate features and statistics
Rank features by discriminating power
Interpret / refine result
Human ImageStream
X
X
X
X
X
X
™
User-defined polarization features
Polarization results in concentration of actin in
a smaller area
Uniform
Polarized
™
User-defined polarization features
Area_Fill-Threshold_Actin (AreaFT50) = area of the filled 50%
threshold mask of the actin image
Uniform
AreaFT50 = 162.5
Polarized
AreaFT50 = 35.0
™
User-defined polarization features
• Risks for the single, user-defined feature approach:
• Did I choose the right mask for the area feature?
• Which threshold mask is best?
• Is filling the mask the right thing to do?
• Is there some other feature that is better than area of a
threshold mask?
• Solution: let IDEAS help!
• Make area features with several mask inputs
• Make multiple actin features using common masks
(default, object, morphology)
• Export features to excel for Rd-based feature selection
™
Calculate multiple features
• Feature algorithms can be selected individually or group-selected by category
• Multiple masks can be applied to the selected algorithms
• One image at a time can be selected for feature calculation
™
Rd Mean stats from IDEAS
• Add population stats from the pop stats context menu
• Select ‘polarized’ for stats population and ‘uniform’ for reference population
• Select ‘Rd - Mean’ for the statistic
• Sort features by image and select the FITC actin image features
• Export to excel and sort data by Rd
™
Polarized vs Uniform actin – best feature
Polarization
1.5
1
0.5
0
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
101
106
111
116
121
126
131
136
Discrimination (Rd)
2
Feature
Polarization features (x-axis) ranked by
discrimination power (Rd, y-axis)
™
Polarized vs Uniform actin – best feature
Conclusion: Area_FT70% provides the best
statistical discrimination between polarized and
uniform actin truth sets
™
Truth set #2: Elongated vs Round cells
Round
Elongated
™
Elongated vs Round cells – best feature
Shape
3.0
2.5
2.0
1.5
1.0
0.5
0.0
1
5
9
13
17
21
25
29
33
37
41
45
49
53
57
61
65
69
73
77
81
85
89
93
97
101
105
109
113
117
121
125
129
133
137
141
Discrimination (Rd)
3.5
Feature
Shape features (x-axis) ranked by
discrimination power (Rd, y-axis)
™
Elongated vs Round cells – best
feature
Conclusion: Shape Ratio_OT(M02) provides the
best statistical discrimination between elongated
and round cell truth sets
™
Classification of cells based on
actin distribution
Challenge = how to pick the feature(s) that best discriminate
cells with differences in actin distribution
Task
Design / execute experiment
High speed imaging
Identify cells (“truth”)
Calculate features and statistics
Rank features by discriminating power
Interpret / refine result
ImageStream Data Interpretation
Human ImageStream
X
X
X
X
X
X
™
Validate classification on entire sample
™
Classification of cells based on
actin distribution
Shape Ratio and Area_FT70% provide the best statistical
discrimination for classifying cells based on actin imagery
Task
Design / execute experiment
High speed imaging
Identify cells (“truth”)
Calculate features and statistics
Rank features by discriminating power
Interpret / refine result
Human ImageStream
X
X
X
X
X
X
™
ImageStream Cytometry:
Discriminating cells based on appearance
• Results are backed by statistics
• Defend images (not algorithms)
• Applicable at any ability level and scales with ability
• You learn features / functions relevant to your application
• Protocol development tool
• Publishable
• AGNOSTIC to application and discipline
™
ImageStream Cytometry:
Discriminating cells based on appearance
Cell signaling
Cell death & autophagy
Internalization &
phagocytosis
DNA damage and repair
Surface and intracellular
co-localization
Stem cell biology
Shape change &
chemotaxis
Oceanography
Cell-cell interaction
Cell cycle & mitosis
Microbiology
Parasitology
™
ImageStream Cytometry:
Discriminating cells based on appearance
Thank you very much for your attention
For more information, including how the ImageStream can advance
your research, go to www.amnis.com
™
Image Analysis Made Easy with
ImageStream Cytometry
April 27, 2011
Thaddeus George, Ph.D.
Director of Biology
Amnis Corporation
™