Upgrade to Pro — share decks privately, control downloads, hide ads and more …

MI_normalization

 MI_normalization

University of Michigan tutorial- segmentation, normalization, and phenotyping

Julia Wrobel

April 03, 2023
Tweet

More Decks by Julia Wrobel

Other Decks in Education

Transcript

  1. Statistical Image Processing
    Segmentation, normalization, and cell phenotyping
    1

    View Slide

  2. • Pixel level processing: work with multichannel tiff files directly
    • Operates on pixel intensity values
    • Cell-level processing: work with tabular data after cell segmentation
    • Operates on median or mean intensity values aggregated at cell level
    • Segmentation
    • Pixel-level
    • Normalization
    • Pixel-level or cell-level
    • Phenotyping
    • Pixel-level or cell-level
    Statistical image processing
    2

    View Slide

  3. Cell segmentation
    3

    View Slide

  4. Identifies cells and nuclei in image
    1. Nucleus channel used to identify nucleus
    2. Cell membrane or cytoplasm markers used to draw boundary around
    nucleus
    Segmentation
    4

    View Slide

  5. • Proprietary software
    • inForm by Akoya, Halo automated image analysis software
    • GUI-based, user-friendly
    • No manually segmented data required
    • GUI-based open-source software
    • CellProfiler, ilastik
    • User-friendly for non-computer scientists/statisticians
    • No manually segmented data required*
    • Deep-learning based open-source software
    • Best performance*
    • Need manually segmented data*
    • Hard to adapt without computational expertise
    Segmentation approaches
    5

    View Slide

  6. Segmentation- Mesmer
    6 Greenwald et. al. Nature Biotechnology, 2022.

    View Slide

  7. Cell phenotyping
    7

    View Slide

  8. Cell phenotyping is the process of
    identifying cell types from marker
    expression values
    • Conceptually, goal is to create a
    “cut point” in marker intensity
    values where cells are either
    positive or negative for a marker
    • Can be performed at pixel or cell
    level
    Phenotyping
    8

    View Slide

  9. Cell types by immune markers
    Thanks to Brooke Fridley and Alex Soupir at the Moffitt Cancer Center for this image

    View Slide

  10. • Can be unsupervised, but often requires expert intervention for cell
    annotation and validation
    • Not specific to multiplex imaging
    • Flow and mass cytometry, single-cell RNA-seq
    • Multiplex imaging has unique challenges
    • Segmentation error leads to phenotyping error
    • Hard to differentiate between “marker positive” and “marker negative”
    Phenotyping
    10

    View Slide

  11. • Marker intensity distributions are
    highly right skewed
    • Often no clear bimodality
    • Sensitive to upstream image
    processing
    • Normalization
    • Cell segmentation
    Phenotyping challenges for MI data
    11

    View Slide

  12. • Marker intensity distributions are
    highly right skewed
    • Often no clear bimodality
    • Sensitive to upstream image
    processing
    • Normalization
    • Cell segmentation
    Phenotyping challenges for MI data
    12

    View Slide

  13. • Marker gating
    • Determine cutpoint in marker intensity histogram to designate “marker positive”
    and “marker negative” cells
    • Unsupervised clustering methods
    • Seurat, Phenograph contain built-in software
    • Most developed for other single-cell analysis
    • Semi-supervised
    • inForm/Halo proprietary software use manual gating to guide phenotyping
    • MAUI/CU-Anschutz: Deep-learning based pixel-level
    • Astir-: Deep-learning based cell-level
    Phenotyping approaches
    13

    View Slide

  14. Image normalization
    14

    View Slide

  15. The slide-to-slide problem
    15
    • Tissues placed on a slide, each contains (10s to 100s) of images
    • Several slides are imaged in the same experiment
    • Multiple sources of noise introduced each time: optical effects,
    instrument parameter tuning, different times of staining for antibodies
    • Large batch effects!

    View Slide

  16. Image transformation, normalization, and batch correction are used to
    make the data more appropriate for downstream analysis by removing
    non-biological biases in marker intensity distributions
    • Transformations: log, arcsinh, square root
    • Make data more normally distributed, do not adjust for systematic effects
    • Normalization: adjusts distribution of marker intensities in each slide,
    image, or channel separately to make distributions appear more similar
    • Batch correction: explicitly removes systematic bias using variables that
    account for processing steps
    Normalizing multiplex imaging data

    View Slide

  17. • Normalization makes the same
    tissue appear similar across
    slides
    • Challenging in MI because
    • No gold standard exists
    • Slides differ in their
    composition
    Evaluating methods for normalization and batch correction
    17
    Unnormalized
    Normalized
    MAP06025
    MAP00083
    MAP03361
    Bottom of crypt
    Interior of crypt
    Stroma Top of crypt

    View Slide

  18. • Histograms reveal differences in
    images intensities across subjects
    and slides
    • Harris et. Al 2022 explored
    combinations of batch correction
    and normalizations
    • Developed new framework for
    evaluating quality of normalization
    • R package mxnorm
    Batch effects in multiplex imaging data
    18
    Harmonization method
    Transformation
    /Normalization

    View Slide

  19. • Most spatial-omics data lack
    ground truth for evaluating slide
    effects
    • Harris et. al 2022 established one
    1. Alignment of marker densities
    2. Cell phenotyping discordance
    3. Proportion of variance due to
    slide
    • Mean division is simple and
    worked well
    Batch correction evaluation framework
    19
    Alignment of marker densities

    View Slide

  20. • Most spatial-omics data lack
    ground truth for evaluating slide
    effects
    • Harris et. al 2022 established one
    1. Alignment of marker densities
    2. Cell phenotyping discordance
    3. Proportion of variance due to
    slide
    • Mean division is simple and
    worked well
    Batch correction evaluation framework
    20
    Cell phenotyping discordance

    View Slide

  21. • Most spatial-omics data lack
    ground truth for evaluating slide
    effects
    • Harris et. al 2022 established one
    1. Alignment of marker densities
    2. Cell phenotyping discordance
    3. Proportion of variance due to
    slide
    • Mean division is simple and
    worked well
    Batch correction evaluation framework
    21
    Proportion of variance due to slide

    View Slide

  22. 22
    • Allows users to easily allow
    evaluate normalization
    procedures in their own data
    • Allows default normalization
    options (from Harris et al. 2022)
    or user specified
    • mxnorm can be used to
    evaluate new normalization
    methods in future papers

    View Slide

  23. • Functional markers are not used to define phenotypes
    • Inform cell function and can be present be present across multiple cell types
    • Interested in differences in expression of functional expression across cell
    or patient populations of interest
    • whether abundance of PD-L1 positive cells differ between responders and
    non-responders to an immunotherapy
    • Still should be normalized!
    • For analysis, using continuous valued intensity is better than thresholding
    • Methods for differential expression analysis for ST data can be used
    • Seal et. al. 2022 provides a method to cluster based on marker densities
    Analysis of functional markers
    23

    View Slide