Advancing Federated Learning of Neural Networks for Medical Imaging

点击翻译按钮获取中文摘要

Machine Learning (ML) algorithms, computational methods that learn to detect patterns in data, promise to improve diagnosis and treatment of disease by enabling fast and accurate medical image analysis. State of the art ML methods, Deep Neural Networks (DNNs), are commonly trained to identify patterns using manually labelled data, such as pairs of medical scans and corresponding manally-generated labels that describe what pathology the scans show. Such labelled medical data are limited because annotation by clinicians is expensive. Moreover, aggregating data from clinical centres across the world in one central database is often infeasible due to privacy concerns. As a result, training databases are small and do not capture the real heterogeneity in clinical practice (rare pathologies, different scanners, etc). Consequently, DNNs trained on such limited data do not generalize well, which hinders their adoption in healthcare. This project will develop methods that enable multiple institutions to collaborate and train a single DNN on their data, without the need to centrally aggregate them in one computational node. This framework is known as Federated Learning (FL) of DNNs between multiple computational nodes (institutions). Models trained with FL could potentially generalize better by learning from diverse databases collected across the world. This can lead to powerful and reliable ML tools for improved disease diagnosis and treatment.FL has the potential to become the standard paradigm for large-scale, international studies on ML for healthcare. There are multiple technical challenges, however, hindering its effective use. This project tackles the following:a) Data acquired at different clinical centres have heterogeneous characteristics, such as due to varying patient demographics or acquisition scanners. When FL is performed between databases with such systematic differences, model optimization is suboptimal. This is because common optimization methods assume the data are identically and independently distributed (iid), which is not true in an FL setting. This project will develop optimization algorithms for FL with non-iid data to improve its effectiveness.b) Performance of existing DNNs is unreliable when applied on data that present different characteristics from those used for training. We will investigate how to identify and model factors of variation between databases used for FL (e.g. from different institutions), enabling inference about expected variability after deployment, to improve model generalization.c) Labels are often limited in healthcare, whereas unlabelled data are abundant. FL methods have been primarily designed for learning using labels. This project will develop FL using unlabelled data, enabling any institution to provide their unlabelled data in a collaborative consortium, to allow models capture better the true data heterogeneity across the world.This research is timely and will advance medical image analysis and the field of ML. Value of FL in healthcare has been demonstrated previously, generating great interest, but technical challenges limit its use. Learning from non-iid and unlabelled data are long standing challenges in ML yet to be solved. Hence results by this project are valuable for medical image analysis but also of interest to other domains. This project falls within the EPSRC Medical Imaging research area and the Healthcare Technologies theme. Its ultimate goal is to create effective FL tools to enable the medical imaging community perform collaborative studies and improve disease diagnosis and treatment.This research is conducted at the University of Oxford within the Institute of Biomedical Engineering, in collaboration with the Big-Data Institute. It is facilitated by existing collaborations with Imperial College London and University of Cambridge, and will seek to establish new ones within UK and internationally.