Comprehensive Summary
Shichao Ma, et al. propose an asynchronous and focal federated learning (FL) approach for skin lesion classification. The study aimed to address data scarcity and class imbalance, which appear in common healthcare tools like telemedicine. The researchers performed a simulation experiment using the PAD-UFES-20 dataset, which collected 1,268 skin lesion images from 798 mobile devices. Each device simulated an individual patient with their own data. The distribution of images modeled a scenario of local data scarcity (few images per patient) and class imbalances (most devices only had images from one class of skin lesions). The performance of this system was tested against popular synchronous FL approaches, including FedAvg, FedProx, and FedNova. The findings showed that synchronous FL approaches performed poorly when the data was scarce and imbalanced, with a performance score ranging from 0.57 to 0.67. At the same time, the proposed asynchronous method achieved high performance scores of 0.78-0.89. The authors emphasize that the asynchronous method is more efficient than synchronous FL in environments with scarce data, which is more realistic for modern healthcare tools.
Outcomes and Implications
Emerging medical technologies, like telemedicine, often receive large numbers of individuals who contribute very small and imbalanced datasets. The asynchronous FL approach was used to address the challenges of data scarcity and class imbalance that cause many current synchronous FL methods to fail. The method presented by the authors primarily aids in teledermatology, where AI models are continuously trained to classify skin lesions directly from patients’ own mobile devices, which is both important to identify skin lesions early and also for patient privacy. While the experiment showed promising results, indicating that the model can be trained to identify skin lesions even with limited data, the authors emphasized that improvements, such as addressing poor-quality or unlabeled data, were needed before real-world implementation.