Priorities, Challenges, and Barriers in Implementing Artificial … · 2020. 8. 5. · Diagnostic...
Transcript of Priorities, Challenges, and Barriers in Implementing Artificial … · 2020. 8. 5. · Diagnostic...
-
Priorities, Challenges, and Barriers in Implementing Artificial
Perception/Intelligence in CRC Screening and Surveillance
Michael Byrne
Clinical Professor of Medicine
Division of Gastroenterology
Vancouver General Hospital
University of British Columbia
-
Financial Disclosures
• CEO and shareholder, Satisfai Health
• Founder, ai4gi
• Co-development agreement between ai4gi and Olympus America
-
Current Clinical Problem
Despite improvements to the tools that doctors use, hospitals are experiencing diminishing returns from better imaging, and doctor performance has stagnated
Doctors need help with real-time
disease detection and diagnosisThe human (doctor) eye is
not accurate enough and is
prone to operator fatigue
Current solutions focus on better
resolution, improved lighting
technology, and other techniques
but doctors’ performance does not
radically improve
-
3D High-Resolution Imaging
-
Prepared by Your Name ©2013 Mauna Kea TechnologiesPrepared by Your Name ©2013 Mauna Kea Technologies
Visual Classes Neighborhood Analysis
Similarity Measurements
AI assisted pCLE Images Classification through CNN training + similarity (BOW) approach
-
Convolution Layer(Variety of filters)
SubsamplingLayer
ConvolutionLayer
Subsampling Layer(Pooling)
Fully Connected Layers
Input Frame ModelPrediction
Type 1
Type 2
No Polyp
Unsuitable
-
“Describe the kinds of training data required to generate a usable
AI model that would be able to identify >99% of polyps in real-
time during colonoscopy?”
• All work to date is encouraging but more work needed for real-time
• False positives a problem
• Detection not all or nothing, unlike some aspects of optical biopsy
• What do investigators define as a polyp for their “AI” detection?
• What defines Deep Learning?
-
Polyp Detection---types of training data needed
• Videos, not “perfect”stills
• Concept of “tracking” and then “encoding” rather than “detection” per se
• “Clinical scenario” quality—stool, spasm, blurred frames, instruments
• Metadata---AI can “extract” information we don’t appreciate
• IEE data for sure, but also white light
• ? “Magnification” data---CLE, Endocytoscopy
• Include other imaging, genomic, treatment data---to allow for AI “prediction”
EMRImaging
Diagnostic Pharma Onco-suitePath, Lab, PGx Surgery
-
Recommended withdrawal time for negative colonoscopy
> 6min
Withdrawal time of different endoscopists3.4 - 9.6min
Nearly half endoscopic physician did not meet the standard
Aasma Shaukat, et al. Gastroenterology 2015;1–6
Quality indicators are not often well followed
Confidential
WHY?Lack of supervision and practical tools!
-
3. Monitoring withdraw speed
Highly Confidential
-
Augment
ScreeningPolyp detection
Transform
Diagnostic Polyp pathology
prediction
Reveal
Treatment Planning ESD v EMR v Surgery
GI Endoscopy: From Procedure to Cost Effective Patient Management
-
17
DataScience
ClinicianBuy-in
OrganizationPriorities
-
15
-
Workflow Integration
-
In the NewsJANUARY 10, 2018
Artificial Intelligence Arrives in GI
Potential downsides of implementing AI
• Decision support/second reader? Or primary
reader?
• If primary, need “human in the loop” (HITL) in AI
solutions
• Self driving car analogy---consumer and patient
acceptance issues
• Data Security, Privacy & Traceability
• Threat to practice
• Billing implications/Insurance reimbursement
https://www.gastroendonews.com/aimages/2018/GEN0118_001a_35481_600.jpghttps://www.gastroendonews.com/aimages/2018/GEN0118_001a_35481_600.jpg
-
What is AI Safety?
18
@romanyam [email protected]
CybersecurityAI + = AI Safety & Security
Science and engineering aimed at creating safe and secure machines.
-
There is a glaringdocumentation gap in endoscopy
Imaging Studies Billing DataLab ResultsElectronic Health Record
Video Recordings
Procedure videos are not being recorded
-
Challenges with traditional endoscopic video systems:
Too complex for every day use
Lack of security for hard media formats
Difficult to scale forbig data
-
Deep learning techniques work best when data is diverse, of
high quality, and in massive quantity.
Cloud video capture helps meet these criteria.
-
Intelligent Real-time Image SegmentationTM (IRIS)
-
©2000-2017 Mauna Kea Technologies©2000-2017 Mauna Kea Technologies23
Endoscopy
Probe-based Confocal Laser
Endomicroscopy (pCLE) provides
realtime optical sections of tissues
Biopsies
X 30
Macroscopic analysis
X 1000 Histology
ex vivo microscopic analysis
Cellvizio
Real time
Compatible
Video rate
Functional imaging
1µm resolution
Invisible information brought to light with
pCLE
-
Thank you
24