Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
L’automatisation de la lecture p16/Ki67
Etude N Wentzensen et al
glandular epithelium squamous epithelium
productive infection
transforming infection
latent infection
persistent infection
Inactivation of pRB by HR-HPV E7 results in
marked overexpression of p16
E2FpRB E7 E2FpRB
Promoterp16INK4a Promoterp16INK4a
Persistent HPV infection
LSIL
Colposcopy,biopsy if lesion
M 12
+
Pap smear
-
p16/Ki67
Inca 2016 French guidelines after abnormal Pap smears
Test HR HPV
Negative Positive
Cytology p16 / Ki 6716/18
Positive
Colposcopy
Primary screening HPV with triage strategiesCombinations of different tests ? Interval between tests if negative?
Test HPV 5 years
NegativeNIIL ASC-us LSILPositiveNegative
Follow up? Follow up? Follow up?
HSIL
HPV 16/18 and p16/ki67 DUAL STAIN (DS) TRIAGE IN A LARGE ORGANIZED CANCER SCREENING PROGRAM of w HPV + (n = 3225)
• DS showed better risk stratification for CIN 3+ compared with cytology in HPV+ women irrespective of genotyping with 30% reduction of colposcopies
• Retesting interval in HPV 16/18 negative with negative p16/ki 67 can be safely extended to 3 years
Wentzensen N et al JAMA 2019
AUTOMATED EVALUATION OF p16/Ki-67 DUAL STAIN
Nicolas Wentzensen, MD, PhD, MSClinical Epidemiology UnitClinical Genetics Branch
Division of Cancer Epidemiology and Genetics
Sydney IPV 2018
• Artificial Intelligence
• Automated Microscopy
• Cloud technology
The Goals of “Artificial Intelligence”
12
Global outreachto customers
Higher medicalaccuracy
Automaticquality control
Load balancing(People,
infrastructure)
Higher efficiency
per lab
AI will assist not replace
MD potential without AI
MD potential with AI
Leverage: „Artificial Intelligence“ as an assistent
CYTOREADER Platform Components
Whole-Slide-Imaging (Digital Pathology)
Cloud-based Data access & Diagnostics
“Artificial Intelligence” Deep Learning
1.
2.
3.
Full microscopy in 60-120s per slideWhole-Slide-Imaging
(Digital Pathology)1.
Cooperation with Hamamatsu Photonics, Japan
Quality-controlledSingle Layer Focus for Cytology
Scanning with a single focus layer
Whole-Slide-Imaging (Digital Pathology)1.
Focus quality score [0-100%]
“Artificial Intelligence“deep learning neural network(CNN) for Quality Control
Cooperation with Hamamatsu Photonics, Japan
Local Infrastructure
AI-basedqualitycontrol
Web browser
Case management
AI-assisteddiagnostics
“Artificial Intelligence“deep learning network(CNN)
Cloud-based Data access & Diagnostics2.
“Use it from anywhere“
AI training of neural network = minimization of 3 image classification errors
Max Robustness(= minimize varianceon distorted images)
Max Sensitivity(=minimize false
negatives)
Max Specificity(= minimize false
positives)
“Artificial Intelligence” Deep Learning3.
Tiling in images (1 cell in median)3.1
3.2
“Artificial Intelligence” Deep Learning3.
Tiling in images (1 cell in median)3.1
Rank all tiles
with AI3.2
Comprehensive computation for each tile of a likelihood between 0 and 1 for precancer
“Artificial Intelligence” Deep Learning3.
Tiling in images (1 cell in median)3.1
Rank tiles with AI3.2
Assisted Diagnosis3.3
Automated detection of DS in Thinprep slides
• Deep Learning algorithm is superior to Support Vector Machine
• Equal sensitivity with increased specificity compared to manual evaluation
Evaluation Biopsy Study (CIN3+) Anal Cancer Screening Study (AIN2+)
AUC Sensitivityp-
valueSpecificity
p-
value
Youden’s
indexAUC Sensitivity
p-
valueSpecificity
p-
value
Youden’s
index
Manual 86.8% ref 40.6% ref 0.27 92.8% ref 36.1% ref 0.29
SVM0.73 83.0% 0.7 40.6% 1.0 0.24 0.75 91.3% 1.0 37.4% 0.7 0.29
CNN40.74 86.8% 1.0 45.7% 0.07 0.33 0.77 91.3% 1.0 46.1%
0.000
10.37
Performance of automated DS (Surepath, KPNC)
EvaluationColpo
referral
p-value
(cytology/
manual
DS)
Sensitivity
p-value
(cytology/
manual
DS)
Specificity
p-value
(cytology/
manual DS)
PPV
p-value
(cytology/
manual
DS)
NPV
p-value
(cytology
/ manual
DS)
Colposcopies
per CIN3+
detected
Pap cytology
(188 CIN3+)
1,860
(60.1%)ref
85.8%
(81.2-90.5)ref
41.9%
(40.1-43.7)Ref
10.1%
(8.7-11.5)Ref
97.5%
(96.6-
98.3)Ref 9.9
Manual DS
(197 CIN3+)
1,536
(50.4%)
<0.0001/
ref
90.0%
(86.0-93.9)
0.2 /
Ref
52.6%
(50.8-54.5)
<0.0001/
ref
12.6%
(11.0-14.3)
<0.0001/
Ref
98.6%
(98.0-
99.2)
0.02/
Ref7.8
Automated
DS >=2 cells
(192 CIN3+)
1,298
(41.9%)
<0.0001/
<0.0001
88.1%
(82.5-91.7)
0.6 /
0.4
61.5%
(59.7-63.3)
<0.0001/
<0.0001
14.8%
(12.9-16.8)<0.0001/
<0.0001
98.5%
(97.8-
99.0)
0.03/
0.86.8
Automated
DS >=1 cell
(201 CIN3+)
1,741
(56.3%)
0.007/
<0.0001
91.8%
(87.3-95.1)
0.05/
0.5
46.5%
(44.6-48.3)
0.06/
<0.0001
11.5%
(10.1-13.1)
0.002/
<0.0001
98.7%
(97.9-
99.2)
0.01/
0.58.7
ROC analysis of DS automation
Vote
•Do you consider assisted AI in Digital Pathology a threat or opportunity for labs in the next 3 years?
A. Threat
B. Opportunity
C. AI will not play an important role
D. None
A Growing NetworkKaiser Permanente, BerkeleyWalter Kinney, MDTom Lorey, MDKiranjit K. Grewal, Patricia E. Goldhoff, MD, Julie D. Kingery, MDDiane Tokugawa, MDNancy Poitras, Alex Locke M.D.
Albert Einstein College of Medicine, Bronx, NYPhilip Castle
Steinbeis Center forMedical Systems Biology HeidelbergBernd LahrmannNiels Grabe
University of OaklahomaRosemary Zuna, MDJoan Walker, MD
NCINicolas Wentzensen, MDMegan A. ClarkeMark Schiffman
UKHDAlexandra KrauthoffLiam BartelsNiels Grabe
Contact: [email protected]
Cerba HealthcareChristine Bergeron, MD