Machine-learned Image Analysis Models for Classifying Liver Fibrosis Stage from Magnetic Resonance Images
Timothy G. St. Pierre1, Michael J. House1, Ajmal Mian2, Sander Bangma3, Gary Burgess6, Richard A. Standish4, Stephen Casey5, Emma Hornsey7, Peter W. Angus5; 1School of Physics M013, Uni- versity of Western Australia, Crawley, WA, Australia; 2School of Computer Science and Software Engineering, The University of Western Australia, Crawley, WA, Australia; 3Resonance Health Ltd, Claremont, WA, Australia; 4School of Medicine, Deakin Uni- versity, Waurn Ponds, VIC, Australia; 5Liver Transplant Unit, Austin Hospital, Heidelberg, VIC, Australia; 6Pfizer Inc, New York, NY; 7Department of Radiology, Austin Hospital, Heidelberg, VIC, Aus- tralia
The aim of the study was to assess the potential of machine- learned image analysis models to
classify liver fibrosis stage in patients with hepatitis C virus (HCV). Methods: Patients with
liver fibrosis due to chronic HCV infection were recruited from the Victorian Liver Transplant
Unit. All subjects had undergone a liver transplant for HCV cirrhosis at least 12 months prior to
the screening visit, and each subject had also had a liver biopsy in the previous 6 months, with a
fibrosis stage score of METAVIR F1 - F3. Twenty eight subjects were recruited and scanned with MRI
at two visits (visit 1 and visit 2). Visit 2 was planned to occur between 35 and 45 days after
visit 1. The distribution of Ishak stage scores on the re-examined pre-study biopsies was 0 (n=2);
1 (n=5); 2 (n=6); 3 (n=8); 4 (n=2); 5 (n=3); 6 (n=2). A 1.5 T Siemens Avanto MR scanner was used to
acquire axial images of the abdomen. A series of breath- hold spoiled gradient echo images was
acquired before and after a bolus injection of 10 mL or 0.1 mmol/kg (whichever was lower) of
gadoterate (Dotarem®) contrast agent includ- ing acquisitions at 3 minutes and 10 minutes after
injection. Machine learning algorithms were trained on the images from visit 1 to classify subjects
into Ishak stage 2 or below and 3 or above. These algorithms extract rotation invariant features by
quantizing gradients from local neighbourhoods of key loca- tions of the images at different
scales. The most relevant feature statistics were then automatically selected to train Support Vec-
tor Machine (SVM) classifiers. The potential of the algorithms to correctly classify patients was
tested by training the algorithms on 27 of the 28 image sets from visit 1 and testing the resulting
model on the 28th image set. This “leave-one-out” strategy was used to classify all 28 image sets
from visit 1. A model trained on all 28 subjects from visit 1 was then used to classify all of the
image sets obtained at visit 2 in order to assess repeatability of the test. Results: Using the
leave-one-out models trained on 27 of the 28 visit 1 images sets, 27, 28, and 28 out of 28 cases
were correctly classified in visit 1 data for pre-, 3-mins post, and 10 mins post-contrast images.
The model trained from all 28 visit 1 datasets correctly classified 23 out of 27, and 26 out of 28,
and 25 out of 28 cases for the pre-, 3-mins post, and 10 mins post-contrast images from visit 2
(corresponding to sensitivities and specificities of 80% & 92%, 93% & 92%, and 93% & 85%
respectively. Conclusions: Machine-learned image analysis models have the potential to classify
liver fibrosis stage from high contrast MR images of the liver in patients with HCV.
Disclosures:
Timothy G. St. Pierre - Consulting: Resonance Health Ltd; Patent Held/Filed: Resonance Health Ltd;
Stock Shareholder: Resonance Health Ltd
Michael J. House - Consulting: Resonance Health; Patent Held/Filed: Resonance Health
Sander Bangma - Employment: Resonance Health Analysis Services Pty Ltd; Stock Shareholder:
Resonance Health Ltd
Gary Burgess - Employment: Pfizer, Conatus
mmittees or Review Panels: Gilead Sciences, BMS;
Grant/Research Support: Gilead sciences
(20min delay)