Next-Gen Analytics in Digital Pathology Cliff Hoyt, CTO Cambridge Research & Instrumentation April 29, 2010 Seeing life in a new light 1 Digital Pathology Today Acquisition, storage, dissemination, remote viewing are maturing Demonstrated value in teaching, frozen sections, referral services, tumor boards Integration into hospital workflow emerging Image analysis tools are improving consistency and providing viable economics for established IHC tests (Her2, ER, etc.) Value proposition for standard H&E yet to be realized 2
Digital Pathology Tomorrow New Analytics Equipping pathologists with automated tools to handle increasing work loads Increasing accuracy of diagnosis & identifying disease subclasses Measuring drug response, for theranostic feedback Capturing tumor heterogeneity, to guide combinatorial therapies Assessing tumor microenvironment reactivity to assess recurrence risk Goal Reliable analytics in the context of the digital slide, compatible with pathology training, and seamlessly integrated into the emerging digital pathology workflow 3 The Dark Report: MSI and automated image analysis in pathology http://www.darkdaily.com/resources
Trainable image analysis Trainable software allows pathologists to teach the computer to mimic visual perception, to automatically select ROIs Automation helps with work load SW can be trained to find all of the tissue of interest across a whole slide, avoiding ROI selection bias and capturing heterogeniety Consistency once set, scoring thresholds are the same every time. Pathologist keeps key roles as trainer and reviewer 5 Identifing disease subclasses with multiparameter analysis Analytical methods that reveal disease phenotype for targeted therapy Multiplexed IHC and IF labeling Cellular multi-parameter protein expression (receptors, phospho-proteins, mirna, etc.) FISH/CISH Combinations of the above 6
Example assessment of signaling activity in lung cancer Autofluorescence DAPI, With inform Component and perk ps6 perk tumor nuclear and (Allexa ps6 mask segmentation and 647) 555) cytoplasm segmentation 7 Comparing patients distinct phenotypes 8
Classifying patient subtype with cluster analysis Per-cell multi-parameter scoring perk Exemplar breast cancer patients perk pakt ps6 perk ps6 pakt pakt perk ps6 Signals from individual cells are divided into quartiles of ps6, pakt and perk expression, and placed into 64 bins. Number in each bin represented as a heat map pakt ps6 9 Analytics also work with IHC Hematoxylin Ki67 red cell cycle phh3 brown cell cycle CC3 gray antiangiogenesis met. suppressor 10
Example in breast cancer - Her Family of Receptors Example in breast cancer - Her Family of Receptors Expression of ErbB family of proteins in Breast Cancer ErbB1 Membrane + Cytoplasmic = Aggressive Tumor & Poor Prognosis ErbB2 Membrane = Aggressive Tumor & Poor Prognosis ErbB3 Nuclear + Membrane = Adverse Cliniopathological properties ErbB4 Cytoplasmic = Prolonged Survival
Capturing heterogeneity through automation Mosaic color BMP Capturing heterogeneity through automation Mosaic color BMP with inform sample finding mask and HPF selection boxes
Capturing heterogeneity through automation With RGB cancer representation mask and of nuclei spectral segmentation cube Scatter plot of ER vs PR expression Field #15 3.83% 10 3 17.29% ER - DAB dab 10 2 10 1 Red = cancer mask Green = cancer nuclei Blue = background 78.31% 0.46% 10 0 10 0 10 1 10 2 10 3 vector red PR - Red Capturing heterogeneity through automation RGB representation of spectral cube Scatter plot of ER Field vs #15PR expression 3.83% Field #5 Field #25 0.00% 1.69% 8.35% 17.29% 10 3 10 3 10 3 50.63% ER - DAB dab 10 2 10 1 78.31% 0.46% 10 0 10 0 10 1 10 2 10 3 vector red PR - Red ER - DAB dab 10 2 10 1 43.74% 54.00% 10 0 10 0 10 1 10 2 10 3 vector red PR - Red ER - DAB dab 10 2 10 1 33.67% 6.33% 10 0 10 0 10 1 10 2 10 3 PR vector - Red red 17.2 3.9 78.2 0.5 0.1 1.8 54.2 43.8 8.4 50.6 33.7 6.3 Mostly Double Negative Significant PR Single Positive Significant Double Positive Multi-parameter, flow cytometry-like data, while maintaining architectural context and heterogeneity. (These fields are from the same sample!) 16
Tumor microenvironment stromal reactivity assessment RGB Image Composite RGB Image Stroma to Tumor Area Ratio 63% % Collagen Area in Stroma 32% Avg Collagen Birefringence in Stroma 9.13 Stem cell density in stroma 175 Segmented DAB Segmented Retardance Image analytics with trainable auto-roi Simple learn-byexample interface require no custom programming. Fast algorithm training Classifiers are reliable Remarkably fast processing (seconds). Efficient and Laptop-compatible inform 18
Automated Multispectral Slide Analysis (Vectra TM ) Fluorescence, brightfield, and multispectral Tissue microarrays and whole tissue sections 200-slide capacity Implements CRi pattern recognition, to significantly reduce scan time and memory usage Bundled with inform software Tissue microarrays Standard Histopathology Fluorescence 19 High speed platforms with high quality digitization 3DHistech Pannoramic whole slide scanner family Multimodal: Brightfield & fluorescence Scalable: slide capacity from 1 to 250 Reliable: advanced industrial designs
Reporting that fits into DP workflow 21 Adding diagnostic value to digital pathology Mosaic color BMP % Area of Stroma 63.12% % Area of Collagen in Stroma 32.48% Avg Retardance of Stroma 9.13 Avg Retardance of Collagen 17.89 Cells/Area (Megapixels) 175.55 www.cri-inc.com