International Symposium for Tissue Phenomics 2016 Translating Tissue Phenomics and Big Data Distillation to Clinical Histopathology Dr Peter Caie Senior Research Fellow QUAntitative and Digital (QUAD) Pathology University of St Andrews Honorary NHS Fife and Lothian
Big-Data Molecular Pathology Genomics e.g. Ion torrent or Illumina Whole genome sequencing Transcriptomics Personalised Big data Pathology e.g. Nanostring or Illumina 800 target multiplex Proteomics e.g. MALDI in situ mass spectrometry or RPPA
Big-Data Histopathology Whole slide imaging Single object co-registration & spatial resolution TMA imaging Hierarchical prognostic or predictive patient stratifiable signatures From population statistics toward personalised histopathology
Fully Integrated Digital Pathology Workflow Zeiss Axio Scan.z1 Leica SCN400 Brightfield & IF Brightfield Remote image access Image Analysis Remote data storage Definiens server workstation Batch processing Licence files & software Teaching Collaborations Remote reporting
Image analysis can objectively quantify histopathological features
Histopathological feature quantification Composite image Standardised objective feature quantification co-registered on single tissue section Multi-plex: Hoechst, PanCK, D240, CD105 PanCK Object mask Composite image D240 Object mask Composite image LVI object mask Composite image CD105 Object mask Composite image BVI object mask Infiltrative subpopulations Vasculature Vasculature invasion Caie PD et al. Journal of Translational Medicine 2014, 12 :156 (1 June 2014) Histopathological features which may prove obscure in H&E stained tissue sections
Muscle invasive bladder cancer Integrate co-registered image analysis based parameters with pathology report: All organ confined MIBC patients between 2007-13 NHS Lothian Integrated Model (BVI) HR=14.2, 95%CI: 3.8-53.1; p<0.0001 BVI LVI LVD BVD TB LTB Age Sex Smoking Grade Stage LVI (TURBT) LVI (cystectomy) Whole slide automated image analysis. Co-registered on single tissue section Clinical pathology report BVI Image analysis
Muscle invasive bladder cancer Advantage of objective quantification through image analysis Kohen s Kappa = 0.11 p= 0.71 (2 consultant pathologists) Inter-rater agreement Pathologist 1: HR=2.3, 95%CI: 0.6-7.9; p=0.2 Pathologist 2: HR= 0.6, 95%CI: 0.1-2.2; p=0.4 Automated BVI HR=14.2 95%CI: 3.8-53.1; p<0.0001
Image analysis can identify novel histopathological features
Proportion of group Proportion of group Proportion of group Proportion of group Validating sub-stratification of CRC Stage II Patients Mean LVD Minimal LVI Whole slide automated image analysis. Co-registered on single tissue section Low LVD High LVD Low LVI High LVI LVI LVD TB LTB (PDC) Composite image Object mask FDR corrected P= 0.558 HR =1.39; 95% CI, 0.46-4.16 Survival (months) Total Tumour budding Cut-off 0.7 vessel percentage of stroma FDR corrected p = 0.05 HR =2.46; 95% CI, 1-6.05 Survival (months) Total PDC Cut-off: 16 LVI events Validation cohort (N = 134). Consecutive stage II CRC patients treated in NHS Lothian over 2 years (2002-3). Low TB High TB Low PDC High PDC FDR corrected P = 0.04 HR=2.49; 95% CI, 1.03-5.99 Cut-off: 253 TB events FDR corrected P = 0.022 HR = 3.2; 95% CI, 1.2 9 Cut-off: 20 PDC events Survival (months) Survival (months)
Discovery of a Novel Prognostic Feature Identify novel histopathological feature through big-data image analysis (Tissue Phenomics): Multi-parametric phenotypic fingerprint from Multi-parametric image signature segmentation shape intensity texture PCA Image analysis Subpopulation big data Dimension reduction CART Model 1 parameter Novel prognostic feature Sum Area too big buds Specificity 86.7% Sensitivity 95% Area under ROC curve 0.90 Validation Random Forest & Gini score Parameter reduction 36 parameters
Novel prognostic feature discovery Spatial relationships of objects and distance maps
Hierarchical Big Data Histopathology Hierarchical image analysis approach: Big data histopathology which captures complex tissue heterogeneity Quantifies candidate histopathological features Exports co-registered multi-parametric data captured in an unbiased manner Decision tree modelling to compare features and identify optimal parameters
Novel prognostic feature discovery Tissue Phenomics Pipeline A. Multiparametric Image segmentation B. Data Handling Subpopulation big data export and collation C. Dimension reduction Phenotypic fingerprint visualisation D. Parameter reduction Random Forest 37 parameters Image analysis Classification And Regression Tree (CART) Model Sum Area PDC Stage II Subpopulation CRC big data Full training cohort Stage II subpopulation Area 161k pxls Good prognosis Random Forest & Gini score Area 161k pxls Poor prognosis E. Prognostic feature identification CART analysis
Integrative pathology signature All colorectal cancer stage II patients (2002-3 NHS Lothian) Significant parameters from the minimum CRC core data set + image analysis algorithm pt stage Area poorly differentiated clusters Differentiation Novel Prognostic Index: HR = 7.5; 95% CI, 3 18.5, p = 0.00001 Clinical gold standard pt stage: HR = 4.3; 95% CI, 1.8 10.3, p =0.0009
Tissue Phenomics in Chondrosarcoma Tissue Studio Tissue Phenomics Benign? Equivocal? Malignant? Mean Area (µm²) Mean Roundness Mean Compactness Mean Shape index Mean Mean Layer 1 Mean Mean Layer 2 Mean Mean Layer 3 Mean Length/Width Mean Length (µm) Mean Number of pixels Mean Density Mean Elliptic Fit Mean Hematoxylin Intensity Mean Optical Density Mean Width (µm) Mean Circularity Mean Ellipticity Mean Ratio Layer 1 Mean Ratio Layer 2 Mean Ratio Layer 3 Std. Dev. Area (µm²) Std. Dev. Roundness Std. Dev. Compactne Std. Dev. Shape inde Std. Dev. Mean Layer Std. Dev. Mean Layer Std. Dev. Mean Layer Std. Dev. Length/Wid Std. Dev. Length (µm Std. Dev. Number of Std. Dev. Density Std. Dev. Elliptic Fit Std. Dev. Hematoxyli Std. Dev. Optical Den Std. Dev. Width (µm) Std. Dev. Circularity Std. Dev. Ellipticity Std. Dev. Ratio Layer Std. Dev. Ratio Layer Std. Dev. Ratio Layer
Tissue Phenomics in Chondrosarcoma Trained machine learning model on cancer cases & benign cases 2 of 15 cases of equivocal have returned with confirmed cancer (model predicted cancer for both) Model Actual Full sample num Cancer Bg 2
PD1&PDL1 spatial resolution - MIBC TS Composer TS Cell simulation & Developer Green panck Blue PD1 Red PDL1 Developer Developer
PD1&PDL1 spatial resolution - MIBC Composite image Developer analysis mask Green panck Blue PD1 Red PDL1 Infiltrative/pushing border, Number/% +ve cells, mean distance maps, proximity, co-localisation, contact, Spatial statistics applied to single object x,y export
PD1&PDL1 spatial resolution - MIBC Composite image Developer analysis mask Green panck Blue PD1 Red PDL1
Green panck Blue PD1 Red PDL1 PD1&PDL1 spatial resolution - MIBC
Integrative multiscale pathology- MIBC Integrate co-registered image analysis based parameters with pathology report: BVI PD1/PDL1 tumour PD1/PDL1 immune LVI LVD BVD TB LTB Age Sex Smoking Grade Stage LVI (TURBT) LVI (cystectomy) Whole slide automated image analysis. Co-registered on single tissue section Clinical pathology report Immune PD1+PDL1 & TB HR=15.4 95%CI: 4-73.1; p<0.0001
Urine cytology bladder cancer screen High grade cancer Bladder cancer: Common recurrence Bi-/Annual cystoscopy Morbidity and discomfort Expensive procedure Urine screen: Non invasive Cheaper Low grade cancer Negative cancer
Urine cytology bladder cancer screen %Total Biomarker cells %Total Biomarker+panCK cells
Low grade cancer High grade cancer High grade cancer Biomarker+ immune cells & panck+biomarker- cells Blue = bladder cancer negative Red = cancer patient
Cytology Phenomics bladder cancer screen % Biomarker+ panck cells Training set of n=50 mean nuclear shape parameter
Quality control steps in Developer Overcoming patient heterogeneity & artefact when applying standardised rulesets to large patient cohorts A) Composite image DAPI channel nuclear segmentation DAPI channel nuclear mask B) DAPI channel nuclear segmentation DAPI channel nuclear mask C) DAPI channel nuclear segmentation nuclear mask Automatic nucleus clean up
Quality control steps in Developer Automatic correction of falsely classified panck objects Autofluorescence Automatic correction of non-specific antibody binding at edge of tissue Composite image Algorithm mask
Quality control steps in Developer Hair? Hair? negated A) False positive stromal cells marked as tumour cells within tumour bud Incorrect cell outline Urinary infection Bacteria negated Composite image Tissue Studio Algorithm mask B) False positive stromal cells negated full cell outline Biomarker +ve punctate Biomarker +ve punctate negated Composite image Developer XD Algorithm mask
Multi-scale omic personalised big data Systems Pathology Big Data Pathology Tissue Phenomics purification ROI Heterogeneous Supopulations & invasive pattern Subcellular expression Individual nuclei Sum values Mean values Ratio values Intensity Std Texture Heterogeneity Morphometrics % positive expression Spatial resolution interaction Co-registration 800-plex proteomics/transcriptomics Whole genome sequencing Clinical pathology report
Acknowledgments Prof. David Harrison (NHS Lothian & St Andrews) Frances Rae Craig Marhsall (NHS Lothian) Dr Chris Bellamy Mustafa Elshani (NHS Lothian & University of Edinburgh) In hwa Um (St Andrews) Graeme Reid (NHS Lothian) Anca Oniscu (NHS Lothian & St Andrews) Hannah Williams (St Andrews & NEQAS) Grant Stewart (NHS Lothian and University of Cambridge) Steve Leung (NHS Fife) Björn Reiß, Jan Gilbert David Harrison, Alex Lubbock & Ian Overton Clinical utility calculator http://www.psycho-oncology.info/cui.html
Image analysis - Chromogen High grade SN45
Image analysis - Chromogen Low grade SN44
pt stage Leibovich risk
3.692e-07