Asynchronous Decentralized Federated Lifelong Learning for Landmark Localization in Medical Imaging

Federated learning is a recent development in the machine learning area thatallows a system of devices to train on one or more tasks without sharing theirdata to a single location or device. However, this framework still requires acentralized global model to consolidate individual models into one, and thedevices train synchronously, which both can be potential bottlenecks for usingfederated learning. In this paper, we propose a novel method of asynchronousdecentralized federated lifelong learning (ADFLL) method that inherits themerits of federated learning and can train on multiple tasks simultaneouslywithout the need for a central node or synchronous training. Thus, overcomingthe potential drawbacks of conventional federated learning. We demonstrateexcellent performance on the brain tumor segmentation (BRATS) dataset forlocalizing the left ventricle on multiple image sequences and imageorientation. Our framework allows agents to achieve the best performance with amean distance error of 7.81, better than the conventional all-knowing agent'smean distance error of 11.78, and significantly (p=0.01) better than aconventional lifelong learning agent with a distance error of 15.17 after eightrounds of training. In addition, all ADFLL agents have comparable or betterperformance than a conventional LL agent. In conclusion, we developed an ADFLLframework with excellent performance and speed-up compared to conventional RLagents.