IVADO 22
Chronic liver disease (CLD) is one of the top ten leading causes of death in the United States and accounts for more than 1,000,000 outpatient and 325,000 emergency department visits.1 Among liver diseases, nonalcoholic fatty liver disease (NAFLD) is recognized as the most prevalent in Western countries, affecting 20-30% of the general adult population, and as much as 90% of diabetic or obese patients.2 Nonalcoholic steatohepatitis (NASH), an advanced form of NAFLD characterized by inflammation occurs in 25% of cases (2.7% of adults in our society)3 and is likely to evolve to fibrosis, cirrhosis and liver failure. This suggests an extrapolated 630,000 NASH cases in Canada. Simple steatosis (fat) is associated with a two-fold risk of developing type 2 diabetes and an increased mortality related to liver and cardiovascular diseases.2,4 NASH is associated with a liver-specific mortality of 15.4 per 1,000 and overall mortality of 25.6 per 1,000 person-years. NASH cirrhosis requires surveillance for hepatocellular carcinoma and portal hypertension, whereas liver failure warrants liver transplantation. NASH has become the second leading etiology of liver transplantation in North America and is predicted to become the leading indication in the near future.5 Therefore, the ability to classify disease severity is crucial for proper patient care. While some ultrasound imaging features visible to radiologists indicate the severity of CLD, evaluation is based on image interpretation.6 Furthermore, inflammation and fibrosis severity cannot be assessed reliably based on ultrasound B-mode images only. Multimodality ultrasound imaging (e.g., elastography) can improve fibrosis staging but challenges remain to assess the whole spectrum of NASH.7
We have recently developed a convolutional neural network (CNNs) based on ultrasound B-mode examinations performed at the Centre hospitalier de l’Université de Montréal (CHUM) for noninvasive detection and stratification of liver steatosis, inflammation, and fibrosis. However, to determine its applicability in clinical practice, we must assess the generalizability of our results and the impact of patient population; imaging views from different positions; scanner manufacturer; and model on classification accuracy. Furthermore, to facilitate development of more robust and general models we aim to integrate our approach into a simulated federated learning paradigm, facilitating learning a model over a collectively larger amount of data.
The primary objective of this study is to simulate federated learning methods to improve the accuracy of chronic liver disease classification in ultrasound imaging. The secondary objectives are to compare the performance of convolutional neural networks (CNNs) trained at one center (CHUM) versus federated learning across partner institutions, to investigate the effectiveness of active learning, domain prior, and self-supervised learning for improving classification accuracy.
The long-term objective of the study is to develop imaging-based biomarkers of chronic liver disease and investigation strategies that may reduce the need for liver biopsy in the future.