Novel Image Analysis Algorithms for Quantifying Expression of Nuclear Proteins assessed by Immunohistochemistry

1
Novel Image Analysis Algorithms for Quantifying
Expression of Nuclear Proteins assessed by
Immunohistochemistry
Elton Rexhepaj , MSc UCD School of Biomolecular
and Biomedical Science UCD Conway Institute,
University College Dublin, Ireland. elton.rexhepa
j_at_ucd.ie
2
Biomarker Validation Application of Tissue
Microarrays
BIOMARKER DEVELOPMENT
3
Interpretation of IHC
Manual
Automated

• Subjective, time consuming
• Inherent intra-observer variability
• Semi-quantitative data
• Pathologist-based analysis remains the current
  standard

• Objective quantification of IHC staining
• Reproducible data
• Continuous output
• A new tool in the hand of the pathologist

4
Application of Image Analysis Approaches to
assess IHC

• Altered nuclear-cytoplasmic ratio of survivin is
  a prognostic indicator in breast cancer
• Automated quantification of ER/PR expression in
  breast cancer patients

Brennan et al resubmitted, Clinical cancer
research, 2007
Rexhepaj et al, manuscript in preparation
5
Altered Nuclear-Cytoplasmic Ratio of Survivin is
a Prognostic Indicator in Breast Cancer

• Promising tumour marker
• Located in the cytoplasm and the nucleus
• Nuclear and cytoplasmic fractions of survivin
  have different biological roles
• Manual interpretation of survivin is challenging
• Conflicting data exists on its prognostic impact
  in breast cancer
• Need for new automated scoring models
• Can automated scores lead to discovery of new
  prognostic
• subgroups

6
Automated image analysis of survivin

• Breast Cancer TMA
• 102 patients
• 0.6mm cores arrayed in duplicate
• Full clinicopathological data
• Median follow-up 77 months
• Image acquistion
• Aperio Scanscope CS Autoscanner

Brennan et al submitted 2007
7
Separating nuclear from cytoplasmic stain
Positive pixel count algorithm
Cytoplasm
HIGH
Cytoplasm nuclear
LOW
We were able to separate cytoplasm from nuclear
staining and independently quantify the IHC
staining intensity
8
Random Forest Clustering Survivin cytoplasmic to
nuclear ratio

• By applying RFC we were able to find 4 cluster
  of patients
• Cytoplasm to nuclear ratio was differently
  expressed in each cluster

Brennan et al submitted 2007
9
CNR and patient survival
High CNR
High CNR
Low CNR
Low CNR

• Clusters with high CNR showed a increase of both
  BCS and OS survival

Brennan et al submitted 2007
10
Cox Regression Analysis of OS
Univariate and Multivariate analysis revealed
that the CNR was a significant predictor of OS in
this cohort along with tumour size and nodal
status
Brennan et al submitted 2007
11
Low CNR a new prognostic subgroup
CytoplasmicNuclear Ratio lt5 (n 78) CytoplasmicNuclear Ratio gt5 (n 18) P value
CytoplasmicNuclear Ratio lt5 (n 78) CytoplasmicNuclear Ratio gt5 (n 18) P value
Tumor Size
Median (Range) 22(10-100) 24 (11-60)
0-20mm 33 (42) 6 (33) 0.6014
gt21mm 45 (58) 12 (67)
ER status1
ER - 26(33) 1 (6) 0.0195
ER 52(67) 17 (94)
PR status1
PR - 35 (45) 3 (17) 0.0335
PR 43 (55) 15 (83)
NHG
NHG I II 37 (47) 18 (100)
NHG III 41 (53) 0
p53 Status2
p53 - 53 (68) 17 (94) 0.0055
p53 25 (42) 1 (6)
Myc Amplification3
Low 43 (55) 17(94) 0.0165
Intermediate/High 16 (21) 1 (6)
Missing 22
A low Survivin CNR is associated with a
mitotic/proliferative phenotype
12
Survivin - conclusions

• Image analysis applied to Survivin IHC
• Image analysis of IHC can produce new automated
  quantitative scoring models
• RFC was used to identify new prognostic subgroups
• Previously unidentified prognostic subgroups can
  be uncovered
• A low Survivin CNR is associated with a
  mitotic/proliferative phenotype

Brennan et al submitted 2007
13
What can be improved
MACHINE LEARNING
MANUAL CALIBRATION

• The supervised approach
• not reproducible and cant be extended to other
  tissue types
• requires a domain expert for the selection of
  validation and test cohort of patients
• The manual calibration
• It is time consuming
• Need to be repeated for each new
  slide/cohort/type of tissue

PATTERN

• Size
• Shape
• Distance
• . . .

Apply the learned or calibrated patterns to the
rest of the cohort.
Alternative Application of non-supervised
learning algorithms to learn the patterns in a
case by case basis
14
Automated image analysis of ER and PR

• Members of the nuclear hormone family
• Expressed in around 70 of breast cancer cases
• Estrogen often induces a multiplication of
  progesterone receptors
• Currently, hormone receptor status is manually
  assessed by a pathologist
• an arbitrary cut off of 10 positive cells is
  used to decide whether a patient should have
  adjuvant hormonal therapy

15
Data
COHORT I

• - 564 pre-menopausal women with primary breast
  cancer
• Patients were randomly assigned to either two
  years of adjuvant tamoxifen

COHORT II
- 512 consecutive breast cancer cases
COHORT III
- 179 consecutive cases of invasive breast
cancer

• more then 1000 patients
• full clinico-pathological follow up

16
Application of IHC nuclear algorithm on tissue
cores examples
17
Algorithm validation
Manual pathologist assessment
Automated percentage
- Validation set -18 representative tissue cores
stained with ER - A trained pathologist was ask
to blindly score each tissue core - A very good
correlation was observed between manual and
automated score
18
Correlation of manual with automated score of ER

• A good correlation was seen between manual and
  automated scores

19
Correlation of manual with automated score of PR

• A good correlation was seen between manual and
  automated scores

20
Selection of the threshold for ER status cohort
I
0.05

• 358 thresholds were generated in the range 0-100
• For each cut-off
• BCS and OS of ER negative patients was compared
  to that of ER positive patients
• The best cut-off for ER was 5 and for 7 for PR

21
A novel approach to automatically define the
threshold for ER status cohort I
- Random forest clustering was used to
automatically cluster patient in ER/- subgroups
22
A novel approach to automatically define the
threshold for PR status cohort I
- Random forest clustering was used to
automatically cluster patient in PR/- subgroups.
23
ER/PR status as defined by clusters and
correlation with manual scores cohort I

• ER status as defined by RFC was correlated with
  manual scores.
• Spearman correlation coefficient was 0.8 for ER
  and 0.7 for PR

24
Correlation of RFC clusters with tamoxifen
response cohort I

• There was a significant effect of 2 years
  tamoxifen treatment on the ER and PR cohort of
  patients as determined by RFC
• No treatment effect was evident in ER-, PR-
  patients as determined by RFC

25

Summary

• - A novel non-supervised image analysis
  algorithm
• - Excellent correlation of manual with automated
  scoring
• - Univariate analysis of OS showed no
  significant difference in the HRs between manual
  and automated scores
• A patient clustering approach to investigate
  patient stratification.
• A new automated approach to stratify patients in
  ER-/
• The ability to predict tamoxifen response was
  similar in manual and automated

26
Acknowledgements
Supervisor UCD School of Medicine and Medical
Science Prof. William Gallagher Dr Amanda
McCann Dr Dermot Leahy UCD School of
Medicine and Medical Science Dr. Donal Brennan
Gallagher Lab Dept of Pathology
Lund University Sweden Dr. Darran OConnor Prof
Goran Landberg Dr. Linda Whelan Dr Karin
Jirstrom Dr. Annette Byrne Asa Kronblad Dr.
Mairin Rafferty Dr. Richard Talbot Dr. Shauna
Hegarty Dr. Helen Cooney Caroline
Currid Sharon McGee Elaine McSherry TARP
Laboratory NCI, NIH, Washington Liam
Faller Dr Stephen Hewitt Ian
Miller Denise Ryan Fiona Lanegan Ben
Collins Tom Lau Karen Power Stephen
Madden Aperio Sarah Penny Aisling O
Riordan Dr Catherine Kelly Dr Sallyann
OBrien
27
EMBO practical course on TissueMicroarray
construction and image analysis
http//coursewiki.embo.org/doku.php?idtissue_micr
oarraysmicroarray_course
June 2008 THE RETURN !!!