Mitigating domain shift in skin cancer recognition with squeeze-excitation attention and dropout-consistent federated averaging

Abstract

Skin lesion classifiers can achieve high in-domain accuracy yet fail in clinical deployment, where hospitals di!er in imaging devices, protocols, and class distributions and patient data cannot be centralized. This thesis first reproduces three state of the art baselines (RegNetY-32GF, VGG16, and SkinLesNet) on three widely used datasets: HAM10000, ISIC (9-class), and PAD-UFES-20. We then propose DermaNet, an E”cientNet-B3 based model augmented with squeeze–excitation attention and trained with focal loss, realistic augmentation, and structured dropout. In centralized training, DermaNet achieves 87.83% on HAM10000, while performance on ISIC and PAD-UFES-20 remains substantially lower due to domain shift. To quantify real world generalization, we evaluate cross-dataset transfer and observe catastrophic collapse. For example, DermaNet trained on HAM10000 drops to approximately 40% on ISIC and 20% on PAD, while the ISIC-trained DermaNet falls to approximately 14% on HAM. We address privacy and heterogeneity via federated learning across three clients and benchmark all frameworks under a FedAvg baseline (FedSE), which improves robustness relative to cross domain centralized deployment (84.54% / 69.55% / 65.29% on HAM / ISIC / PAD). However, FedAvg with stochastic dropout su!ers from zero dilution. Structurally dropped weights are averaged as true zeros, weakening attention and slowing convergence. Finally, we propose FedMD, a mask-aware aggregation strategy that averages only active (non-dropped) parameters using deterministic client masks, yielding consistent gains over standard FedAvg, On HAM it achieved 85% accuracy, on ISIC it achieved 71% accuracy and in case of PAD-UFES it achieved 75% accuracy.