Using Cloud-Based Deep Learning AI Platform to Analyze Gigantic - - PowerPoint PPT Presentation

Using Cloud-Based Deep Learning AI Platform to Analyze Gigantic Pathology Images

Kaisa Helminen, CEO Fimmic Oy October 11th, 2017

Every third person affected 14 Million new patients in 2012 +50% more by 2030

Cancer

Increasing number of samples

Problem

Lack of pathologists Increasing number of samples

Gap!

Subjective analysis Manual visual methods Risk for variability in diagnosis

Microscopy in tissues diagnostics

Digitalization

• >

Deep Learning AI

Easy sample archiving and retrieval Sharing, remote consultation Machine vision & Deep Learning AI -assisted analysis, e.g. % of tumor tissue, tumor grading, identification of infection, quantification of certain features, etc.

Digitalization enables

How to create a virtual slide?

Images captured at high magnification Up to 100 000 image tiles Stitched digitally and compressed to a large picture montage (Gb - Tb)

Samples

WebMicroscope Workflow

Artificial Intelligence

Pathologists Researchers Any microscope scanner Educators Any device

Artificial Intelligence & Deep Learning

Facial recognition

id-labs.org The Guardian

Self-driving cars Tissue diagnostics

Different types of Deep Learning Image Analysis tasks

• 1. Laborious quantification tasks, combined with region of interest

selection, e.g. quantification of certain cells in epithelium

• 2. Segmentation of tissue based on morphology, e.g. tumor grading,

epithelium/stroma segmentation

• 3. Detecting and quantifying rare targets, e.g. infectious agents, forensic

samples

Training of deep learning classifiers

1

Original Labelled

Whole slide samples

2

Training set from sample regions

3 4

Deep Learning Application to new samples

Epithelium

Epithelium segmentation from breast cancer samples.

Application example 1 - Quantification task

Breast cancer diagnostics, Quantification of Ki67+ cells from epithelium

• 1. Epithelium-stroma

segmentation

• 2. Quantification of Ki67 + and -

cells inside the epithelium

Negative and weak signals Moderate and Strong

Context-intelligent Image Analysis

Enables full automation Removes extra staining step

• > Saves time

Accurate Reproducible

• > Supports correct diagnosis

Application example 2 - Automated segmentation of tumor area

Prostate cancer, Segmentation of cancer tissue, area quantification

H&E stained prostate tissue Result image - segmentation

Application example 3 - Quantification task

Testicular Cancer, Quantification of tumor infiltrating lymphocytes (TIL%)

H&E of immune cell rich region Heat map showing immune cell detec7on

FIMM – Oxford collabora0on 2017, unpublished results

Digitized whole slide images of testicular cancer are huge gigabyte-sized files Areas of infiltrating immune cells detected by automated analysis includes millions of immune cells (red areas)

Application example 3 - Quantification task

Testicular Cancer, Quantification of tumor infiltrating lymphocytes (TIL%)

FIMM – Oxford collabora0on 2017, unpublished results

Example patient: Immune cells in testicular cancer

Automated counting result Total immune cell count = 768.349 Immune cells/square mm tumor = 4223 Details of the analysis shown in the 
 video

FIMM – Oxford collabora0on 2017, unpublished results

Application example 4 - Quantification Task

Quantification of fat accumulation in liver cells

Consistent Accuracy and Reproducibility over large sample sets

Application example 5 - Quantification from segmented area

Quantification of fibrosis in liver tissue

Significant time savings Reproducibility

Application example 6 - Quantification Task, complex background

Quantification of nerve cell bodies from rat brain tissue (Parkinson’s, Alzheimer’s) Significant time savings:

From 45 minutes to 0,5 minutes analysis

Unforeseen Accuracy & Reproducibility

Application example 7 - Quantification from selected tissue compartments

Quantification of glucagon+ alpha cells from Islets of Langerhans in pancreas

Significant time savings Reproducibility

Application example 8 - Identification of rare targets

Detection of Malaria infection in red blood cells

Examples of performed Deep Learning Image Analysis Applications

All algorithms are intended for Research Use Only.

Breast cancer biomarkers: ER, PR, Ki67 + Epithelium/stroma segmentation Breast cancer, mitosis quantification Prostate cancer: Gland and epithelium segmentation Lung cancer, mouse tissue: Tumor burden, tumor classification Colon cancer, Ulcerative Cholitis Seminoma (testicular cancer): TIL% Liver biopsies: Hepatosteatosis, fibrosis Rat brain: Nerve cell bodies (Parkinsons, ALS research) Forensic pathology: sperm detection from smears Blood: RBC, WBC, Platelets, Malaria parasites etc.

Immunofluorescence Images

WebMicroscope - Intelligent Image Analysis in Cloud

Advanced Image Storage and Collaboration tools in Cloud

Compatibility Efficient compression

Deep Learning Algorithms & Cloud computing

No local hardware Indefinite possibilities for algorithms

Disruptive business model

Affordable SaaS model for all sizes of projects

The Future of Pathology is Digital

Supportive data for decision making -> Prognosis -> Suggesting treatment -> Faster, more accurate diagnosis and cure

Johan Lundin MD, CSO Co-Founder Board Member Mikael Lundin MD, Director of Concept Design Co-Founder Board Member Kari Pitkänen Business Development Co-Founder Board Member Mikael Jääskeläinen Sales Manager Kaisa Helminen CEO Tuomas Ropponen CTO Sartorius Thermo Scientific Finnzymes Sartorius Fisher Scientific Finnzymes FIMM Fisher Scientific Finnzymes, co- founder, sold to Thermo Fisher Scientific in 2010 FIMM Karolinska Institute HUS FIMM University of Helsinki Outotec Biohit Delta-Enterprise Previously: Antti Merivirta Marketing Manager 360Visualizer Testure Finland

Experienced Core Team

Combination of life science entrepreneurs, software development and machine vision experts & recognized scientists.

Contact

Kaisa Helminen, CEO +358 40 679 0669 kaisa.helminen@fimmic.com www.webmicroscope.com @kaisa_helminen