High Content 2019 6 th Annual Conference September 17 th -19 th , Boston, MA Joseph B. Martin Conference Center Educational Program: Intro to HCS/HCA Image and Data Analysis Mark-Anthony Bray, Ph.D Novartis Institutes of BioMedical Research Cambridge, Massachusetts, USA [email protected]
42
Embed
High Content 2019 Annual Conference September 17 , Boston, MA … · 2019-09-18 · High Content 2019 6th Annual Conference September 17th-19th, Boston, MA Joseph B. Martin Conference
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
• Also known as segmentation: Partitioning an image into regions of interest
• Step 1: Distinguish the foreground from the background by picking a good threshold
• Foreground: Regions where I(x,y) > threshold T
9/11/2019 SBI2 High Content 2019 11
Illumination Correction
• Nonuniformities introduced in the optical path of the sample, microscope, and/or camera
9/11/2019 SBI2 High Content 2019 12
• Example: Uneven illumination from left to right – Can lead to inaccurate segmentation and measurements
– Cell at (a) is brighter than (b) even if cells have same amount of fluorescent material
(a) (b)
Carpenter et al, Genome Biology 2006, 7:R100
Illumination Correction
• Recommendations
• Create new illumination correction if switching microscopes
• Perform per-plate correction
• Perform per-channel correction, as absolute illumination intensities may differ between channels9/11/2019 SBI2 High Content 2019 13
Images from Carolina Wahlby
Input image Output image
=÷
Approximation of
backgroundAverage many images
Fit continuous function to result
or smooth heavily
Background Subtraction
• Top-hat (“rolling ball”) filtering
9/11/2019 SBI2 High Content 2019 14
Image Thresholding
What is the best threshold value for
dividing the intensity histogram into
foreground and background pixels?
9/11/2019 SBI2 High Content 2019 15
Raw input
image
Thresholded
binary image
0: Background
1: Objects
Labeled objects
Colored ROI:
Connected
pixels
Here?
Or here?
Pixel values
Fre
qu
en
cy
Pixel-Based Image Classification
• For images where a threshold cannot be found…
• Machine-learning tools can be helpful, e.g., ilastik• User manually labels regions of image• Suite of features are used to distinguish regions and create a classifier
9/11/2019 SBI2 High Content 2019 16
Sommer and Gerlich, JCS 2013, 126:1
Outline
• The image as quantitative data
• Identifying the image foreground
• Splitting object clusters
• Identifying cellular compartments
• Measurement extraction
• Statistical analysis
9/11/2019 SBI2 High Content 2019 17
Separating Touching Objects
• Once the foreground blobs have been identified, what next?• Thresholding is not sufficient to separate clustered or touching objects
9/11/2019 SBI2 High Content 2019 18
• Step 2: Distinguish multiple objects contained in the same foreground blob
Watershed Segmentation
• Consider the image as a surfacewith basins….
9/11/2019 SBI2 High Content 2019 19
••
•
••
•
•
•
http://www.svi.nl/watershed
Images from Carolina Wahlby
Separating Touching Objects
• Identifying objects: Some options
9/11/2019 SBI2 High Content 2019 20
Peaks
2
1 2
Indentations
1
• Intensity-based: Works best if objects are brighter at center, dimmer at edges
• Shape-based: Works best if objects have indentations where objects touch (esp. if objects are round)
1
••
•
••
•
•
•
Outline
• The image as quantitative data
• Identifying the image foreground
• Splitting object clusters
• Identifying cellular compartments
• Measurement extraction
• Statistical analysis
9/11/2019 SBI2 High Content 2019 21
Identifying Cell Objects
• Nuclei more easily separated than cells
• DNA markers are specific
• Yield good foreground/background contrast
• Uniform shape
• Identifying cells is more difficult• Available markers often lower contrast
• Unclear boundaries between cells, depending on the cell type and culture conditions
9/11/2019 SBI2 High Content 2019 22
Secondary Object Identification
• “Growing” the primary objects to identify cell boundaries
• Use segmented nuclei as “seeds” by using a cell stain channel
• Some assays do not require precise cell ID
• E.g, is a protein located in nucleus or cytoplasm?
• Produce proxy cells by growing nuclei by Npixels if no cell stain available
9/11/2019 SBI2 High Content 2019 23
Identifying Subcellular Structures
• With appropriate markers, other
subcellular compartments can be labeled
• These can be identified using the same
methods already mentioned
• Consider using enclosing object as mask
for better pre-processing, thresholding
• Make sure to assign subfeatures to
enclosing objects
9/11/2019 SBI2 High Content 2019 24
Pre-processing
Sub-object ID
Sub-object relation
Outline
• The image as quantitative data
• Identifying the image foreground
• Splitting object clusters
• Identifying cellular compartments
• Measurement extraction
• Statistical analysis
9/11/2019 SBI2 High Content 2019 25
Measuring Object Counts
• Most common readout• # of cells per image/well
• # of organelles per image/well
• # of organelles per cell
• Number of objects per image/well is often a useful readout for QC purposes
9/11/2019 SBI2 High Content 2019 26
Measuring Object Morphology
• Reduce an aspect of object shape to a single value
• Example features
• Area: Pixel coverage of object
• Perimeter: Length of object boundary
• Eccentricity: Object “oblongness”
• Major, minor axis length: Object elongation
• Form factor: Measure of compactness
• Zernike features
• Objects touching the image border should be excluded if shape is important
9/11/2019 SBI2 High Content 2019 27
http://www.perkinelmer.co.uk/
Measuring Object Intensity
• Example features• Integrated (total) intensity: Sum of the object
pixel ∝ amount of substance labeled
• Mean, median, standard deviation intensities
• Lower/upper intensity quartiles
• Correlation coefficients between channels: Colocalization
• Make sure to illumination correct beforehand
9/11/2019 SBI2 High Content 2019 28
• Related to the amount of marker at a pixel location
Images courtesy of Ilya Ravkin
Measuring Object Texture
• Determine whether the staining pattern is smooth or coarse at a particular scale
• Selecting the appropriate texture scale
• Higher scale: Larger patterns of texture
• Smaller scale: More localized (finer) patterns of texture
• Shows all the measurements acquired• For each individual cell • In every image • In the entire experiment.
9/11/2019 SBI2 High Content 2019 36
+1
0
-1
Cell #6111617
-.2 .7 -.1 0 .2 -.9
Data Normalization
• Used to remove systematic errors from the data
• Allows comparison of screening runs from different plates, acquisition times, etc.
• Ideally, results in:• Similar measurement ranges observed across different wells with the same treatment• Similar measurement distributions of the controls (positive or negative)• Keep in mind the recommendations from Assay Development section!
• Common approaches• % of control: Divide by mean of corresponding measurement from control• % of samples: Divide by mean of corresponding measurement from all samples• Z-score, robust Z-score: Transform to zero mean/median, unit variance/MAD
• Alternative approach: Normalized value = percentile within rank-ordered data
9/11/2019 SBI2 High Content 2019 37
Statistical Analysis Software
• Spreadsheets (e.g., Microsoft Excel)• Widely used because of familiarity,
• Unable to handle large screening datasets
• Lack sophisticated analysis methods
• HCS/HTS microscope vendors often bundle data-analysis functionality with hardware, image-analysis software
9/11/2019 SBI2 High Content 2019 38
Statistical Analysis Software
• Specialized commercial tools• Wide variety of products• Often bundled with hardware• Talk to vendors for more details