AI-powered computational pathology for liver diseases€¦ · Month Portal Inflammation Log-rank p
Transcript of AI-powered computational pathology for liver diseases€¦ · Month Portal Inflammation Log-rank p
Proprietary & Confidential
AI-powered computational pathology for liver diseasesOscar Carrasco-Zevallos, PhDSenior Scientific Program ManagerPathAI
Interpretation of liver histology is prone to error and current scoring systems do not fully capture disease heterogeneity
Interpretation of liver histology is prone to error and current scoring systems do not fully capture disease heterogeneity
1 David E. Kleiner et al. “Design and validation of histological scoring system for nonalcoholic fatty liver disease.” Hepatology 20052 Zach Goodman, et al. “Grading and staging systems for inflammation and fibrosis in chronic liver diseases.” Journal of Hepatology 2007
Ishak scoring system CRN NAFLD activity scoring system
Histologic scoring systems have limited reproducibility
1 Beth A. Davison, et al. “Liver biopsies in nonalcoholic steatohepatitis (NASH) clinical trials.” Hepatology 20202 Zach Goodman, et al. “Grading and staging systems for inflammation and fibrosis in chronic liver diseases.” Journal of Hepatology 20073 PathAI Analysis: AASLD 2019, median interval between biopsy re-reads, 16 weeks (range 9, 20).
Intra-observer discordance for grading key NASH features grading is high3
(Particularly for lobular inflammation and ballooning, 22–47% of cases)
Number of Biopsies
Steatosis Lobular Inflammation Ballooning
KappaCases with
Discordance KappaCases with
Discordance KappaCases with
Discordance166 0.69 22% 0.38 42% 0.66 22%
162 0.50 29% 0.29 43% 0.43 36%
149 0.59 26% 0.42 39% 0.29 47%
◆ Published literature has shown only moderate levels of inter- and intra-reader concordance for grading key features of chronic hepatitis and NASH
• Inter-reader kappas were 0.61, 0.48, 0.33, and 0.52 for steatosis, fibrosis, lobular inflammation, and ballooning, respectively1
• Inter-reader kappas were 0.4-0.6 for portal inflammation, interface hepatitis and parenchymal injury and inflammation2
Proprietary & Confidential
AI-powered pathology for HBV and NASH◆Machine learning (ML) models trained
to interpret liver histology with 100% reproducibility
◆Designed for rigorous quantification of key histologic features
◆ Elucidate associations of ML histologic features with disease progression, clinical outcomes and response to therapy
ML model development for automated assessment and quantitation of liver histopathology
Pokkalla H, et al., Oral presentation #187, AASLD 2019Juyal D, et al., Poster presentation LBP31, EASL ILC 2020 (Late-breaker abstract submission)
ML model development for automated assessment and quantitation of liver histopathology
Pokkalla H, et al., Oral presentation #187, AASLD 2019Juyal D, et al., Poster presentation LBP31, EASL ILC 2020 (Late-breaker abstract submission)
Tissue regionsPortal inflammationLobular inflammationInterface hepatitisBallooningSteatosisMicrovesicular steatosisBile duct
ML-based quantification of histologic features of chronic inflammation correlate with expert pathologist assessment1
Juyal D, et al., Poster presentation LBP31, EASL ILC 2020 (Late-breaker presentation)1: Pathologist reads followed clinical trial protocol
Portal inflammationInterface hepatitis
Ishak periportal necrosis grade
ML
inte
rface
hep
atiti
s ar
ea, %
M
L po
rtal i
nfla
mm
atio
n ar
ea, %
Ishak portal inflammation grade
ML-based quantification of histologic features of NASH correlates with consensus pathologist assessment1
Pokkalla H, et al., Oral presentation #187, AASLD 20191: Consensus pathologist reads obtained for research purposes
SteatosisBallooningLob inflam
Pathologist consensus grade
ML
stea
tosi
s ar
ea, %
10
0
20
40
30
50
3n=4
0 1 2n=22 n=101 n=37
r=0.86; p <0.001
Mod
el S
core
,%
Consensusmedian grade
1
3
2
4 r=0.47; p <0.001
Mod
el S
core
,%
00 1 2 3
n=25 n=80 n=47 n=12Consensus
median grade
STEATOSIS LOBULAR INFLAMMTION
ML
lob
infla
mar
ea, %
ML
ballo
onin
g ar
ea, %
Pathologist consensus gradePathologist consensus grade
50 r=0.86; p <0.001
10
0
20
40
30
50
3n=4
0 1 2n=22 n=101 n=37
r=0.86; p <0.001
Mod
el S
core
,%
Consensusmedian grade
1
3
2
4 r=0.47; p <0.001
Mod
el S
core
,%
00 1 2 3
n=25 n=80 n=47 n=12Consensus
median grade
STEATOSIS LOBULAR INFLAMMTION
F0 F1 F2 F3 F4 F5 F6
ML Ishak fibrosis patterns
ML fibrosis score quantifies heterogeneity and correlates with expert pathologist assessment1
0 1 2 3 4 5 60
2
4
6
Ishak Stage by Central Reader
Slid
e-le
vel M
L Is
hak
Scor
e 𝜌=0.60p <0.001n=456
F0 F1 F2 F3 F4 F5 F6
F0 0.8%F2 3%
F314%
F418%
F552%
F612%
Manual Ishak Stage 6 ML Ishak Score 4.54
Juyal D, et al., Poster presentation LBP31, EASL ILC 2020 (Late-breaker presentation)1: Pathologist reads followed clinical trial protocol
In subjects whose cirrhosis regressed at year 5, fibrosis improvement by year 1 is evident only on ML Ishak score
n=22 n=8 n=22 n=8 n=22 n=80
2
4
6
ML
Isha
k Sc
ore
p=0.039 p=0.043p=0.146
Isha
k St
age
byC
entra
l Rea
der
p=0.235 p<0.001p=0.156
p<0.001
p=0.120
No cirrhosis at Year 5Cirrhosis at Year 5
◆22/30 HBV subjects achieved cirrhosis regression after 5 years of therapy
◆Subjects who achieved cirrhosis regression at year 5 had significant reduction in ML Ishak score from baseline to year 1
◆Fibrosis improvement was not evident on manual histology by year 1
Juyal D, et al., Poster presentation LBP31, EASL ILC 2020 (Late-breaker presentation)
Baseline Year 1 Year 5
ML-based histologic features are predictive of progression to cirrhosis in subjects with bridging fibrosis due to NASH
Pokkalla H, et al., Poster presentation #2497, EASL ILC 2020 (Selected for “Best of ILC Summary”)
Surv
ival
free
of
dise
ase
prog
ress
ion
,%
Baseline ML CRN fibrosis score
113/755 of subjects had progression to cirrhosis
Baseline ML hepatocellular ballooning area, %
Surv
ival
free
of
dise
ase
prog
ress
ion
,%
◆ Progression to cirrhosis was associated with higher ML CRN fibrosis score at baseline (HR 2.66 [95% CI: 1.82, 3.90])
◆ Progression to cirrhosis was associated with higher ML ballooning proportionate area at baseline (HR 1.87 [95% CI: 1.20, 2.91])
Ratio of portal/lobular inflammation is associated with increased risk of clinical disease progression in NASH
Pokkalla H, et al., Oral presentation #187, AASLD 2019
No ClinicalEvent n=653
ClinicalEvent n=21
0
50
100
150
200
0
20
40
60
80
100
Even
t-Fre
e Su
rviv
al,%
p <0.001
Rat
io o
f por
tal t
o lo
bula
rin
flam
mat
ion
0 3 6 9 12 15 18 21 24Month
Portal Inflammation
Log-rank p <0.001Hazard ratio 4.50(95% CI 1.86, 10.85)
Event-free Survival According to Ratio of Portal
to Lobular Inflammation
<40
≥40
Boxes depict median (IQR); whiskers based on Tukey method.Richardson MM, et al. Gastroenterology 2007;133:80-90; Gadd VI, et al. Hepatology 2014;59:1393-1405; Brunt EM, et al. Hepatology 2019;70:522-31.
Proprietary & Confidential
Conclusions◆ PathAI ML models enabled reproducible
and quantitative assessment of liver histology beyond that afforded by manual scoring
◆ In research studies, ML read-outs:
• Revealed treatment-associated histologic improvement not evident by manual scoring
• Were predictive of disease progression and liver-related clinical events
Acknowledgments
We extend our thanks to the patients, their families, and all participating investigators.
These studies were funded by Gilead Sciences, Inc.
Gilead SciencesVithika Suri
Anuj GaggarJohn F. Flaherty
G. Mani SubramanianLing Han
Catherine JiaRyan S. Huss
Chuhan ChungRobert P. Myers
PathAIDinkar Juyal
Chinmay ShuklaHarsha Pokkalla
Zahil ShanisQuang Le
Maryam PouryahyaAmaro Taylor-Weiner
Benjamin GlassKishalve PethiaMurray Resnick
Michael MontaltoIlan WapinskiAditya KhoslaAndrew Beck
CollaboratorsIra JacobsonEdward Gane
Maria ButiStephen A. HarrisonZachary Goodman
Arun J. SanyalZobair M. Younossi
Thank You
Oscar Carrasco-Zevallos, PhDemail: [email protected]