Kuchinke Personalized Medicine tools for clinical research networks
Simagis for healthcare
1. SMART IMAGING TECHNOLOGIES web-pathology.net
PERSONALIZED CANCER THERAPY
INTEGRATION OF KNOWLEDGE
Digital Pathology and Machine Learning
for Healthcare Organizations
2. Personalized Cancer Therapy: Knowledge Path
Personalized
Therapy
Medical
History and
Personal
Information
Genetic
Information
Pathology
Information
(tumor
biomarkers)
Personalized cancer therapy is a treatment strategy centered on
the ability to predict which patients are more likely to respond to
specific cancer therapies.
This approach is founded upon the idea that tumor biomarkers
are associated with patient prognosis and tumor response to
therapy.
In addition, patient genetic factors can be associated with drug
metabolism, drug response and drug toxicity.
Personalized tumor molecular profiles, tumor disease site and
other patient characteristics are then potentially used for
determining optimum individualized therapy options.
Source: MD Anderson Cancer Center
Pathology, the “study of disease”,
is an essential component for
analysis of personalized cancer
therapy options
PERSONALIZED CANCER THERAPY
3. Utilizing Pathology Knowledge: The Challenge
Traditionally, pathology diagnosis is presented in descriptive
natural language statements. Often It is verbose professional
opinion of human expert with little quantitative information
Statistical agreement between human experts is 75%-85%.
Pathology diagnosis is rendered by pathologist observing
patterns of cells on tissue slide under the microscope. In order
to be useful for comparison and analysis these observations
must be:
• Quantified
• Objectified
This can be achieved (in theory) by digitizing pathology slide
and applying image analysis algorithms to quantify cell pattern
expressions.
PERSONALIZED CANCER THERAPY
4. Analyzing Pathology Slides: Machine Learning
Machine Learning Neural Networks learn to recognize images in the
same way humans do – by example, rather than by formalized
“handcrafted features”.
Since 2012 major improvement in visual recognition was achieved
with so called deep learning neural networks. Latest generation of
Visual Recognition Neural Networks achieve accuracy of recognition of
natural objects similar to human observers. This area of technology is
experiencing explosive growth.
Using Machine Learning brings number of advantages to visual
recognition applications:
• No need to formalize complex “handcrafted features”, pathologist
can just point to patterns they need to recognize
• No dependency on image analysis engineers (almost)
• System can be trained on very large number of samples to achieve
robust recognition
• New data samples can be added to model easily to increase
accuracy
PERSONALIZED CANCER THERAPY
5. Diagnostic Pattern Library: Applications and Benefits
Classified cancer pattern library is valuable digital asset that can
be licensed to other parties to train visual recognition and image
analysis algorithms.
Visual recognition application can be used to automatically
annotate digital pathology slides and link them with the rest of
institutional cancer knowledge base. This application can be
licensed to third parties to use for the same purposes.
Research and Clinical Applications:
• Computer-assisted cancer diagnosis with pre-screening,
suggestive diagnosis options and contextual links to cancer
knowledge libraries (similar cases, experts, research, additional
tests etc.)
• Data mining and of advanced analytics of historic tissue
samples for cancer patients with known outcomes with
purpose of building predictive knowledge bases for cancer
care and drug discovery
PERSONALIZED CANCER THERAPY
6. Machine Learning: Requirements
Machine Learning approach to pattern recognition creates
new functional requirements for digital pathology software
• Robust visual recognition models need large number of training images which
requires more time for annotating than single pathologist can provide. This problem
can be solved by utilizing number of pathologists creating annotations for training
Collaborative Training (Crowdsourcing)
• Digital Pathology system should have capability for extracting specially formatted
image data sets on demand for training neural networks
Training Data Extraction
• Training of Neural Network requires massive parallel GPU computing power for a
short time. This scalable computing power can be economically delivered by scalable
cloud infrastructures such as Amazon AWS.
Cloud Deployment
• Digital pathology software should be able to send image areas to neural network
application for recognition and visualize responses for user.
API Integration and Visualization Interface
PERSONALIZED CANCER THERAPY
7. Our Solution: Pattern Recognition with Machine Learning
• Last generation deep learning convolution networks can identify target tissue patters with 95%
accuracy
Deep Learning
• Pathologists can train recognition solution by simply annotating target tissue patters on slides
in their workspace
• They can easily set up classes of patterns for identification
Easy Training
• Robust solutions can be trained from multiple slides to identify target tissue patterns reliably
across large variety of samples
• Recognition models can be retrained easily if new patterns or different samples should be
added
Robust Recognition Models
• Slides in digital archives can be processed automatically for pattern detection and labeled
based on findings
• New slides can be analyzed and classified on upload with suggestive classification available
when human expert opens the slide
Automatic Processing
• Visualization overlays help quickly locate and review target patterns
• Visualization layer provides quantitative information about patterns
Advanced Visualization
• All data is stored in the database and available for search, data mining and analytics
Powerful Analytics
Our software can train neural networks and utilize latest deep learning visual recognition solutions
from best in class solution providers
PERSONALIZED CANCER THERAPY
8. Pathologist are located in different places.
Patients may get care at different locations
Accountable Care Model
International Opportunities
Organizations need technology to:
• Complete diagnostic workflow seamlessly between different remote locations
• Work with different scanners that use different file formats
• View slides fast from any location, keep pathologists productive
• Access relevant information for slides without LIS
• Distribute work to different people, manage workflow and access
• Share work and collaborate remotely
• Catalog and search information and results
• Integrate pathology slides with other patient information
Integration of Pathology Knowledge: Drivers and Needs
INTERGRATION OF KNOWLEDGE
9. Distributed Workflow
• Slides from multiple remote scanners are automatically
uploaded to designated workspaces on central cloud server with
user friendly web interface
Pathologist-centric Architecture
• Slides from any scanner are shown in consistent diagnostic view
with slide labels and overview
Support for all Scanners
• Application Interface is accessible from any browser / OS with
no plugins to install in the browser
“Pure” Web Interface
• System delivers complete case information to pathologists
including document attachments
Full Case Information
Our software platform supports seamless diagnostic
workflow across geographic boundaries and IT networks.
INTERGRATION OF KNOWLEDGE
10. Pathologist Productivity
• Pathologist workspace is an integrated “cockpit”, the single place with all tools for
completing diagnostic workflow
• Case includes all information needed to render diagnosis
• The information fields in workspace are customizable for specific workflow and case use
Integrated Cockpit
• Multiple individual and team workspaces facilitate distribution and transition of cases
between team members
Work Spaces
• Configurable notifications alert user when new work arrives
Notifications
• In a few clicks pathologists can create and distribute professional media-rich reports with
original images
• Report templates are customizable for specific workflow and case use
Web Reports
• Pathologists can use “canned” text to quickly enter repeatable information
Text Templates
• Flexible tags and attributes help organize information for quick reference and navigation
Tags and Attributes
Our software provides features and interface that make
pathologist productive with digital pathology workflow
INTERGRATION OF KNOWLEDGE
11. Collaboration
• Multiple users can work on the same case at the same time
Multi-Access
• Sharing of slide, case or entire workspace is possible via simple URL
• Sharing can be done securely with authorized users only or with larger
groups with simple pre-shared links
Easy Sharing
• Real-time annotations provide rich information instantly visible to
multiple viewers at the same time
Rich Annotations
• Report distribution to group of recipients can be done with a single
click
• Reports include case data, annotations and original images
• PDF or HML format options available
• Report distribution is HIPPA compliant
Report Distribution
Our digital pathology solution opens opportunities for
productive team-based pathology not possible with glass slides
INTERGRATION OF KNOWLEDGE
12. Integration: Data Mining and Discovery
• Non SQL flexible indexed database architecture allows integrated
storage of different data items across multiple locations
Distributed Database
• Flexible structure allows storing and integrating various data in the
single information store
• New data can be added to database structure at any time
Comprehensive Data
• Selection and navigation is possible for any data item in the database
• Global search on any data is instant even for millions of items
Instant Search and Navigation
• Data items can be linked with external data sources and knowledge
bases such as diagnostic codes, SNOMED classifications or
proprietary knowledge bases
Data Linking
We provide instant search, navigation and data mining ability across millions of slides
INTERGRATION OF KNOWLEDGE
13. Integration: Information Systems
• RESTful API with live examples and templates provide easy
integration with third-party applications
Easy API
• Integration with other medical information systems is available
via HL7 Integration Engine (Rhapsody by Orion Health)
LIS / EMR Systems
• Third party image analysis application can access images and
metadata
Algorithms
• Information in the database can be integrated with other web
based knowledge system via standard integration protocols
Knowledge Bases
Our product includes standard industry data exchange protocols and
APIs for integration with any third party application
INTERGRATION OF KNOWLEDGE