EUVIMEDEUVIMED entdecken
Crossref

FED DERM-SLA INVESTIGATING CONFLICT-AWARE SEMANTIC KNOWLEDGE TRANSFER FOR FEDERATED SKIN LESION CLASSIFICATION UNDER LABEL-SPACE AND DOMAIN HETEROGENEITY

Acta Innovations · 30.10.2026 · 64 (1s)

DOI: 10.62441/actainnovations.vi.2837 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Acta Innovations, online erschienen am 30.10.2026.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Federated learning (FL) enables collaborative medical image analysis while keeping patient images decentralized. However, heterogeneity in diagnostic label spaces and image distributions presents substantial challenges for effective knowledge sharing across participating institutions. In particular, transferring class-level semantic information between clients may introduce negative transfer when local feature representations are insufficiently compatible. This study investigates Fed Derm-SLA, a federated skin lesion classification framework integrating a shared ResNet-18 feature encoder, client-specific classification heads, semantic class prototypes, and similarity-dependent knowledge transfer. Three controlled configurations are examined: federated learning without semantic transfer (A0), ungated semantic transfer (A1), and similarity-gated semantic transfer (A2). Experiments are conducted using 10,015 HAM10000 dermoscopic images partitioned into four heterogeneous clients with lesion-disjoint training, validation, and testing sets. Model performance is assessed across three matched random seeds and further evaluated using 1,077 curated ISIC 2019 images. On the internal test set, A0, A1, and A2 achieve mean macro-F1 scores of 0.6682, 0.6590, and 0.6665, respectively. Gated transfer consistently improves internal macro-F1 relative to ungated transfer across all three seeds, although it does not exceed the no-transfer baseline. External macro-F1 decreases to 0.3698, 0.3658, and 0.3625, respectively, indicating substantial generalization challenges. Transfer diagnostics also indicate that the similarity gate accepts all observed candidate transfers, providing little evidence for effective conflict rejection. These results suggest that semantic transfer weighing alone does not necessarily lead to better classification or cross-dataset generalization in the heterogeneous federated setting. The study offers a controlled empirical evaluation of prototype-based transfer and points out significant limitations for the design of more discriminative conflict-aware federated learning mechanisms.

Zitieren

APA

(2026). FED DERM-SLA INVESTIGATING CONFLICT-AWARE SEMANTIC KNOWLEDGE TRANSFER FOR FEDERATED SKIN LESION CLASSIFICATION UNDER LABEL-SPACE AND DOMAIN HETEROGENEITY. Acta Innovations, 64(1s). https://doi.org/10.62441/actainnovations.vi.2837

Vancouver

FED DERM-SLA INVESTIGATING CONFLICT-AWARE SEMANTIC KNOWLEDGE TRANSFER FOR FEDERATED SKIN LESION CLASSIFICATION UNDER LABEL-SPACE AND DOMAIN HETEROGENEITY. Acta Innovations. 2026;64(1s). doi:10.62441/actainnovations.vi.2837

BibTeX
@article{euvimed2026,
  title = {FED DERM-SLA INVESTIGATING CONFLICT-AWARE SEMANTIC KNOWLEDGE TRANSFER FOR FEDERATED SKIN LESION CLASSIFICATION UNDER LABEL-SPACE AND DOMAIN HETEROGENEITY},
  journal = {Acta Innovations},
  year = {2026},
  volume = {64},
  number = {1s},
  doi = {10.62441/actainnovations.vi.2837}
}
Kostenlos starten →