- 现金
- 62111 元
- 精华
- 26
- 帖子
- 30437
- 注册时间
- 2009-10-5
- 最后登录
- 2022-12-28
|
LBP31
Machine learning identifies histologic features associated with
regression of cirrhosis in treatment for chronic hepatitis B
Dinkar Juyal1, Chinmay Shukla2, Harsha Pokkalla2, Amaro Taylor2,
Oscar Zevallos2, Murray Resnick2, Michael Montalto2, Andrew Beck2,
Ilan Wapinski2, Patrick Marcellin3, John F. Flaherty4, Vithika Suri4,
Anuj Gaggar4, Mani Subramanian4, Ira Jacobson5, Edward Gane6,
Maria Buti7. 1PathAI, Boston, MA, USA; 2PathAI, Boston, MA, USA;
3Department of Hepatology, AP-HP Hôpital Beaujon, Clichy, France;
4Gilead Sciences Inc., Foster City, CA, USA; 5Department of Medicine, NYU
Langone Health, New York, New York; 6Auckland Clinical Studies,
Auckland, New Zealand; 7Hospital General Universitario Valle Hebron
and Ciberehd, Barcelona
Email: [email protected].
Background and aims: Machine learning (ML) may facilitate
interpretation of histologic changes associated with treatment of
chronic hepatitis B virus (HBV) infection. We developed ML modelsthat identify and quantify histologic features in a clinical study of HBV
patients receiving antiviral therapy.
Method: ML models were developed using H&E histology images
from 330 patients enrolled in registrational studies for tenofovir
disoproxil fumarate for HBV (GS-US-174-0102, GS-US-174-0103).
Histological improvement and regression of cirrhosis were assessed
by a central pathologist at baseline (BL) and weeks 48 and 240
according to the Ishak/Knodell necroinflammatory scoring and Ishak
fibrosis staging systems. Images were split into training (N = 1090)
and testing sets (N = 1061). Models were trained using the PathAI
research platform (Boston, MA) to identify inflamed regions and
immune cells (lymphocytes and plasma cells) using annotations from
40 board-certified pathologists. Additional annotations of steatosis
and ballooning from previous models were included in training.
Slide-level, quantitative ML features were computed and correlated
with pathologist scores to assess accuracy. Regression analysis was
performed to determine associations of ML features at BL and
changes from BL with cirrhosis regression at week 240. Results were
generated using the testing image set.
Results: ML % area of portal inflammation correlated strongly with
Ishak portal inflammation scores (ρ = 0.643; p < 0.001) and ML % area
of interface inflammation correlated strongly with Ishak periportal
necrosis scores (ρ = 0.716; p < 0.001). Of the 48 patients in the testing
set with cirrhosis at BL, 36 patients (75%) no longer had cirrhosis at
week 240. Lower ML % area of steatosis (Figure) and greater ML % area
of lobular inflammation at BL were predictive of cirrhosis regression
(p = 0.006, p = 0.047, respectively), indicating the presence of underlying
fatty liver in those who do not resolve cirrhosis. Change from
baseline in ML % area of portal and lobular inflammation as well as
change in lymphocyte density correlated with regression of cirrhosis
at week 240 (p = 0.010, p = 0.026, p = 0.031 respectively).
Figure: Steatosis area at baseline, determined by machine learning, was
predictive of cirrhosis regression at week 240.
Conclusion: An ML approach accurately classified histopathologic
features in H&E images from HBV clinical trial biopsies. ML features at
BL and changes in ML features with treatment were significant
associated with cirrhosis regression. An ML approach for evaluating
liver histology in patients with HBV can provide mechanistic insight
into both HBV disease pathogenesis and cirrhosis regression. |
|