Clinicians, Leveraging AI expertise, Understanding the
Regulatory Framework
• Clinical interpretability guiding the adoption of AI-first
differential diagnoses
• Disparity in data literacy affecting the communication of
AI among healthcare scientists
• Regulatory challenges impeding the penetration of AI
into clinical practice - from research to policymakers
Frank J. Rybicki MD, PhD, Professor and Chair, Department
of Radiology, UNIVERSITY OF OTTAWA FACULTY OF
MEDICINE
Packaging the Monolith - PHP Tek 2024 (Breaking it down one bite at a time)
The Patient Journey with AI
1. FRANK J. RYBICKI, MD, PhD
The patient journey with AI:
saves time, saves money, saves lives
Professor and Chair, Dept of Radiology | University of Ottawa Faculty of Medicine
Chair, Appropriateness Criteria® | American College of Radiology
2. DISCLOSURE
Frank J. Rybicki
is the Medical Director of
Imagia Cybernetics
M I C H E L A N G E L O
“Rondanini Pietà”
Milan, Italy
3. REVOLUTIONS IN THE PRACTICE OF RADIOLOGY
Hardware Software
Canada Science and Technology Museum
Ottawa, ON, Canada
Change HealthCare (formerly McKesson)
6. LUNG CANCER CARE
Overview
Radiology Radiology Pathologist Oncologist/ Radiation OncologistPulmonologist / Surgeon
Mature Risk
stratification Guide Biopsy / Inform Pathology Personalized Treatment & Response Prediction
Screening Diagnosis Treatment
Imaging Imaging (Staging) Biopsy Molecular Diagnosis
Detection Specificity Pathology Precision
1 2 3 4
L U N G R A D S
Earlier Stages Later Stages
Surgery Medical Therapy
Chemo, Immuno, Targeted
Radiotherapy
M.D.PATIENTA.I.
Quality and Safety Programs
12. ACR LungRADSTM
CT is sensitive with high NPV,
it is not specific
Detection Errors
Errors in perception, analyses,
and categorization
Limited Specificity Essential, but
Limited Data
13. SCREENING WITH AI
Overview
Specificity is enhanced
with more comprehensive
data that undergoes more
complete analyses
Errors can be
minimized with
data and
computational
resources
Mature Risk stratification
Screening / Diagnosis
Imaging Detection + Staging
RadiologyA.I.PATIENTWITHAIM.D.
WITH
A.I.
14. SCREENING DIAGNOSIS
Detects and predicts
Lung CAD detection &
malignancy predictor
Disconnect in performance between
AI science and clinical care
NLCST: n=1000 pts, 16 cancers
AI science: 98% accuracy for detection
Therefore, no guarantee of detection
for any of the 16 cancers. (Detection
is not screening.)
Partnerships: if clinicians do not
define the problem, AI scientists will
not make the most meaningful
contributions.
19. DIAGNOSTIC PITFALLS
Advanced imaging (e.g.
combined PET & CT) is
limited by expense and
expertise, particularly
in some parts of the
world. There are also
technical limitations.
L I M I TAT I O N S
CHEST
X-RAY
Unable to
detect small
lung tumors.
CHEST
CT
Insufficient
to confirm
cancer.
MRI
SCAN
Rarely used
for lung
lesions
because signal
intensity is
very low
PET
CT SCAN
Expensive and
resource is
very limited.
False positive
for infection
(e.g. Tb)
BONE
SCAN
Replaced by
PET CT to
detect lung
cancer that
has spread to
the bones.
Medical Imaging
20. DIAGNOSTIC PITFALLS
Tissue sampling
Multiple biopsy
strategies - each
has its limitations,
but all have:
Risk to patient
High expense
Defined rate of
inadequate tissue
Lung regions that are
difficult to access
Primary & Secondary
Bronchus can be reached
with traditional biopsy tools
Tertiary Bronchus, that
can be reached with
bronchoscopes & EBUS RP
Traditional
Bronchoscopy
EBUS radial
probe (RP)
EBUS radial
probe (RP): Blind
21. Small biopsy samples
Traditional pathology
from small biopsy
samples has
diagnostic limitations
Tumor heterogeneity
Intertumour
heterogeneity
Intratumour
heterogeneity
Intercellular genetic
and non-genetic
heterogeneity
Subclone 3
Subclone 1
Subclone 2
Clonal
heterogeneity
DIAGNOSTIC PITFALLS
22. PITFALLS OF DIAGNOSES
1
EXPENSIVE AND
LIMITED RESOURCES
Medical Imaging
Advanced imaging (e.g.
combined PET & CT) is
expensive and in some
parts of the world (e.g.
Canada & Europe) the
resources are very limited
2
GENERAL
LIMITATIONS
Multiple biopsy strategies
- each has its limitations,
but all have:
Risk to patient
High expense
Defined rate of
inadequate tissue
3
DIAGNOSTIC
LIMITATIONS
Traditional pathology from
small biopsy samples has
diagnostic limitations
Tumor heterogeneity
23. IMPROVED STRATEGIES FOR DIAGNOSIS
Tumor heat map to
guide biopsy Averting repeat biopsies
Informed request for tissue
Pathology needs
Immunology needs
Oncology needs
28. Medically, no one was
“wrong”. But why did
the system arrive at the
diagnosis so slowly, and
so miserably?
Patient
Professionals
System
29. THE LUNG CANCER PATIENT: PITFALLS OF THERAPY
Tumor Board is an
essential
component of
oncology
therapies
All board rooms are not equally shared
Personalized treatment based on
side-effect/outcome
Delayed Physician communication /
Siloed Expertise
Best intent is the norm, not the
exception. However, decisions are
difficult and at times arbitrary
30. AI MODIFIES THE PATIENT JOURNEY
Mature Risk
Stratification Guide Biopsy / Inform Pathology Personalized Treatment & Response Prediction
Screening Diagnosis Treatment
Imaging Imaging +/- Biopsy +/- Molecular Diagnosis
Detection Specificity + Staging Pathology Precision
1 2 3 4
L U N G R A D S
Early Stage Late Stage
Surgery Medical Therapy
Chemo, Immuno, Targeted
Radiotherapy
Screening / Diagnosis
Imaging Detection + Staging Augment
Molecular Tests
Inform Biopsy
Treatment
RadiotherapySurgery Medical Therapy
Predict Recurrence
& Side Effects
Dose Optimization
Treatment
Options
Personalized
Regimen
Radiotherapy
Radiology Pulmonology / Surgery Oncology
PATIENTA.I.PATIENTWITHAI
Quality and Safety Programs
M.D.
WITH
A.I.
31. GUIDING TREATMENT FOR VARIOUS STAGES
Determine Surgery vs SRS
based on recurrence and distant metastasis risk
Patient 1 – Lower stage tumor
Surgery vs radiation vs early combined Tx
Some guidelines - Safety
AI driven treatment safety profile
Patient 2 – Higher stage tumor
Aggressive conventional therapies
Target therapies to be considered - Efficacy
AI driven informed decision on therapies
32. SUMMARY AND REFLECTIONS
The next revolution needs integration
Infrastructure and reimbursement poses
significant friction to accelerated integration
Best practices are now obligated to incorporate AI
Fear being surpassed by enthusiasm
It is OK to have fewer physicians
Just ask leadership
33. SUMMARY AND REFLECTIONS
I read lung cancer screening CT scans
Lung cancer was one of many examples
Screening, diagnosis, treatment, <monitoring>
Tumor board of the future
Software with a seat at the table
No other way to handle the information
Siloed approach to medical data is flawed