Asynchronous and focal federated learning for skin lesion classification under local data scarcity and class imbalance

Background and Objectives: Federated learning (FL) is an approach that enables the training of machine learning (ML) models using data from multiple data nodes without direct data transfer, hence making it a good choice for healthcare ML applications to alleviate data privacy and security concerns. Most standard FL approaches focus on the setting of a small number of nodes, with each node contributing a sizable amount of data. However, in emerging healthcare settings such as telemedicine and the Internet of Medical Things (IoMT), it is necessary to consider the situation in which there is a large number of nodes, and each contributes a relatively small number (data scarcity) of non-independent (class imbalance) data points.

Methods: In this paper, we propose an asynchronous and focal update approach to enable FL to address this problem. In particular, we demonstrate its use in a teledermatology setting, in which a skin lesion image classifier is continuously updated based on data in a highly distributed network of mobile devices. We performed a situation experiment in which 1,268 skin lesion images across 798 mobile devices contributed to the training of a 3-class classifier in an FL framework.

Results: We found that widely used synchronous FL methods perform poorly under conditions of data scarcity and imbalance. Specifically, using FedAvg, FedProx, and FedNova, the trained classifiers achieved AUROC values of 0.57-0.67, 0.63-0.66, and 0.64-0.67, respectively, on the held-out test set across various experimental settings. In contrast, our proposed asynchronous and focal approach achieved a test AUROC of 0.78-0.89 after 40 global training epochs. This performance is significantly closer to the optimal AUROC of 0.91, which is achievable by training a classifier with all the data on a centralised server without FL.

Conclusions: These results demonstrate that our approach provides a useful solution to implement an efficient FL scheme under the conditions of data scarcity and class imbalance that are commonly found in realistic telemedicine and IoMT applications.