RePFL: lag-resilient personalized federated learning for skin cancer detection under extreme non-IID heterogeneity

Early and accurate detection of skin malignancies is critical for improving patient outcomes. However, deep learning (DL) models often fail to generalize because of variations in imaging protocols, differences in lesion appearance, and the uneven distribution of clinical data across institutions. Although federated learning (FL) enables collaborative model training without sharing raw patient data, its practical deployment is hindered by non-identically distributed (non-IID) data and inconsistent participation from slow or unreliable clients (laggards). We propose RePFL, a lag-resilient personalized federated learning (PFL) framework for skin cancer classification. RePFL integrates late-stage feature embedding alignment to reduce representation divergence while preserving client-specific personalization, a computationally efficient stochastic trace-based Fisher Information Matrix (t-FIM) score to estimate the contribution of each client without constructing the full Fisher matrix, and semi-asynchronous staleness-aware aggregation to incorporate delayed client updates while reducing the impact of lagging clients. Experiments on the HAM10000 dataset demonstrate that RePFL achieves classification accuracies of 87.50% and 86.86% under two non-iid settings, outperforming established FL and PFL baselines. The framework preserves raw-data locality but does not provide formal privacy guarantees, and its explanation maps are evaluated using computational faithfulness and stability metrics rather than clinical validation. Overall, RePFL provides a practical and robust framework for multi-institutional collaborative learning in resource-constrained healthcare environments.