Federated-Learning-based Hierarchical Diagnosis of Liver Fibrosis

Hepatic fibrosis is an important prognostic factor as severe liver fibrosis may lead to liver cancer or even death. To grade liver fibrosis, ultrasound gray-scale images and ultrasound elastic images are commonly used in clinical diagnosis to judge the severity of liver fibrosis. However, these two diagnoses methods are often vulnerable to disturbances, such as personal experience or instrument differences. Moreover, these individual differences usually lead to conflicting stand-alone machine learning diagnosis models at each hospital whose medical data are not allowed to share in public due to data privacy. To handle the conflicts among diagnosis models, we propose a federated learning based hierarchical diagnosis method of liver fibrosis by utilizing shear wave elasticity pictures of multiple users across hospitals without sharing the original data. Our method is validated with authentic shear wave elasticity pictures of hepatic fibrosis patients in Shanghai, China. Experimental results show that our method is able to preprocess these shear wave elasticity pictures, train local diagnosis models at each hospital and securely consolidate into a shared global diagnosis model whose accuracy is over 70% with only a small dataset containing a few hundreds of labeled pictures. Our method is expected to further improve in its accuracy with more training samples. Our method would be the first practice based on federated learning in liver fibrosis diagnosis.