SchaeferTP105

CD30 Cell Graphs of Hodgkin lymphoma are not scale-free — an Image Analysis Approach

Tim SchäferGoethe-University Frankfurt am Main, Germany

Institute of Computer ScienceDepartment of Molecular Bioinformatics

ISMB 2016, Orlando July 12, 2016

H. Schäfer1,T. Schäfer1, J. Ackermann1, N. Dichter1, C. Döring2, S. Hartmann2, M.-L. Hansmann2, and I. Koch1. Bioinformatics, 32(1):122–129, 2016.

1 Molecular Bioinformatics, Goethe-University Frankfurt, Germany2 Dr. Senckenberg Institute of Pathology, University Hospital Frankfurt, Germany

Pathology workflow at Dr. Senckenberg Institute of Pathology

Biopsy StainingTissue & cell morphology

Suggestion for treatment

Digital pathology

Biopsy

Slide scannerWhole slide images (WSIs)

Digital image analysis

Cell propertiesQuantificationGuidanceData for a model

- Lymphadenitis- Hodgkin lymphoma - NScHL - MCcHL

StainingTissue & cell morphology

Suggestion for treatment

Classical Hodgkin Lymphoma (cHL)

● Malignancy of the lymphatic system, Thomas Hodgkin

● No solid tumor is formed● Hodgkin-Reed-Sternberg (HRS) cells

● Complex tumor microenvironment, shaped by HRS cells


● Malignancy of the lymphatic system, Thomas Hodgkin

● No solid tumor is formed● Hodgkin-Reed-Sternberg (HRS) cells

● Complex tumor microenvironment, shaped by HRS cells

● CD30 marker: HRS cells and activated lymphocytes, e.g., in

– cHL Subtypes:● Nodular sclerosis (NScHL)● Mixed cellularity (MccHL)

– Lymphadenitis (LA)

Blue: Cell nuclei (haematoxilin)Red: CD30+ cells (fuchsine red)


● Pathologists inspect small image regions

● Morphology of some interesting CD30+ cell cells

● Tissue morphology

● Patterns?


Project goals

● Work on the full WSI

● Systematic analysis and quantification of:

● CD30+ cell properties

– Typical cell size, shape, ...● The spatial distribution of CD30+ cells

– Cell density, neighborhood, ...

– Patterns

… in the different diseases and disease sub types.


Project goals

● Long term goals:● Understand how CD30+ cells spread

through the lymph node and lymphatic system

● Find out more about the composition of and the cellular communication in the tumor microenvironment

● Combine data from different projects to build a model


Dataset

● Dr. Senckenberg Institute of Pathology

● 35 whole slide images (WSIs)● Pyramidal image format

● Resolution 0.25 μm per pixel (px)

● Dimension up to 100,000 x 100,000 px

● Uncompressed file size up to 30 GB

Dataset

● Dr. Senckenberg Institute of Pathology

● 35 whole slide images (WSIs)● Pyramidal image format

● Resolution 0.25 μm per pixel (px)

● Dimension up to 100,000 x 100,000 px

● Uncompressed file size up to 30 GB

Lymphadenitis Mixed Cellularity cHL Nodular sclerosis cHL

NS cHL

Input image

NS cHL

Detected cells, colored by morphological properties

The Impro Image Processing Software

● Java Advanced Imaging API● Interfaces to CellProfiler and ImageJ

## Impro PIPELINEImproImageViewerPlugin;CreateImproImage

ImproImageMultiResClusteringPlugin;IdentifyROI:minLayer=3;maxLayer=2;saveROI=true;onlyProcessROI=true;noImageOutput=false

ImproCellProfilerAdapter;SplitImage:tileSize=1024;border=100;useROI=true;createSubfolder=false;outputDir=$WORKSPACE$/tmpDirs/

ImproCellProfilerAdapter;IdentifyCellObjects:pipelineFile=pipeline_CD30.cp;dir=$WORKSPACE$/tmpDirs/;useBatchFile=true;numberOfProcesses=4

MainApp;ExitProgram

Command file (headless mode) Graphical user interface

CellProfiler: Lamprecht et al., Biotechniques, 42:71, 2007.ImageJ: Abramoff et al., Biophotonics International, 11(7):36–42, 2004.

Parallel processing of image tiles

DatabaseCellProfiler: Kamentsky et al., Bioinformatics 27(8): 1179-1180, 2011.

CellProfilerPipeline

Impro

Impro: Detect region of interest

● Images contain one or more tissue patches and background

● We are interested in the tissue area

Impro: Detect region of interest

● Images contain one or more tissue patches and background

● We are interested in the tissue area

● Detect tissue on low-resolution image● Reduces data by ~40% on average

Schäfer, et al. Computational Biology and Chemistry, 46:1–7, 2013.

Imaging Pipeline

1. Input image

Imaging Pipeline

1. Input image

2. Color deconvolution

Color deconvolution method: Ruifrok, Analytical and Quantitative Cytology and Histology, 19(2):107–113, 1997.

Imaging Pipeline

1. Input image


3. Segmentation, Region growing, Shape descriptors


CellProfiler: Kamentsky et al., Bioinformatics 27(8): 1179-1180, 2011.

Imaging Pipeline


CellProfiler: Kamentsky et al., Bioinformatics 27(8): 1179-1180, 2011.

1. Input image


3. Segmentation, Region growing, Shape descriptors

4. Filtered resulting objects

Validation

Manual annotation Software

● For a selection of randomly chosen image tiles:

Validation

+Validation

False positive (FP)

True positive (TP)

False negative (FN)

}

Sensitivity = TP / (TP + FN)Precision = TP / (TP + FP)

WSIs are large and heterogeneous

Blood vessels Very high cell density

Stain residues, broken and folded tissue Stitching error from scanning process

Positions of cells detected by the imaging pipeline, colored by morphological properties

How to model it?

Unit disk graphs

x

y

Unit disk graphs

t

Distance threshold t

Unit disk graphs

t

Unit disk graphs

t

NS cHL

Tissue and detected cells

Tissue and graph

CD30 cell graph:up to 90,000 cells and 7,000,000 edges

Clustering? Typical patterns?

Does this graph contain any information?

● Could it be random?

● What would we expect a random unit disk graph to look like?

● How do the properties of our measured CD30 cell graph differ from a random unit disk graph?

Null hypothesis – Create equivalent random cell graph for an image

● For a single image, distribute the same number of cells on the same area

● Poisson point process:● The position of each cell

is chosen randomly

● Each position is equally probable

● Locations of existing cells do not influence new cells

Vertex degree distribution of the null model: Poisson and simulation

Vertex degree k

p(k)

Null model for a single Nodular sclerosis case

Avg. degree of the case ~ 11

Null model and measured cell graph

Vertex degree k

p(k)

Data for a single Nodular sclerosis case

Gamma distribution fit to cell graph

Vertex degree k

p(k)


CD30+ cells cluster in the tissue

Vertex degree k

p(k)


Cell graph analysis

● Comparison of vertex degree distributions and clustering by disease type revealed differences between LA and the cHL sub types

● Extend analysis● More graph properties

● Integration of data like cell shape

● Additional markers for other cell types

Work in progress: neighborhood analysis by cell shape and size

Jennifer Scheidel,Unpublished results

Work in progress: neighborhood analysis by cell shape and size

Jennifer Scheidel,Unpublished results

● Next neighbor class● Next neighbor distance

Summary

● Imaging pipeline for cell detection in whole slide images● Quantification of cell and image properties (typical cell size, cell counts,...)

● Images and cases are very heterogeneous

● Definition of cell graphs to model cells in space

● Comparison with null model revealed● Cell graphs are not random

● CD30+ cells cluster in the tissue

– Attraction, cell division, lymph node structure?● Vertex degree distribution of cell graphs can be modeled by the Gamma

distribution

● Cell graphs show differences between lymphadenitis and the Hodgkin lymphoma sub types

Acknowledgments

Molecular Bioinformatics GroupGoethe-University Frankfurt am MainIna KochHendrik Schäfer, Jörg Ackermann, Jennifer Scheidel, Patrick Wurzel, Tanmay Pradhan, Marie Hebel, Sonja Scharf, Norbert Dichter

Travel funding to ISMB 2016 was generously provided by IRB-Group.

Acknowledgments

Dr. Senckenberg Institute of PathologyProf. Dr. Dr. h.c. Martin-Leo HansmannProf. Dr. Sylvia HartmannDr. Claudia Döring

Thank you for your attention!

Questions?

Appendix slides follow

MCcHL case

LA case

NScHL case

Color deconvolution

CD30 (Fuchsine red) Haematoxilin

Method: Ruifrok et al., 2004

Heterogeneous images

High intensity Low intensity

Unspecific staining

Documents

SchaeferTP105