Federated Learning for Decentralized Artificial Intelligence in Melanoma Diagnostics.

Area of Science:

• Digital pathology
• Artificial intelligence in healthcare
• Medical diagnostics

Background:

• Developing AI for melanoma classification typically requires large, centralized datasets, raising patient privacy concerns.
• Federated learning (FL) offers a decentralized alternative, distributing model development across institutions without sharing sensitive patient data.

Purpose of the Study:

• To evaluate if a privacy-preserving federated learning approach can achieve diagnostic performance comparable to centralized and ensemble learning methods for AI-based melanoma classification.
• To investigate the utility of FL in multicentric melanoma diagnostics using histopathological whole-slide images.

Main Methods:

• A multicentric, single-arm diagnostic study involving 6 German university hospitals.
• Development of a federated model for melanoma-nevus classification using prospectively acquired whole-slide images.
• Benchmarking the federated model against classical centralized and ensemble approaches using holdout and external test datasets.

Main Results:

• The federated approach showed significantly lower performance than the centralized approach on the holdout dataset (AUROC 0.8579 vs 0.9024).
• However, on the external test dataset, the federated approach (AUROC 0.9126) significantly outperformed the centralized approach (AUROC 0.9045).
• The federated approach performed significantly worse than the ensemble approach on both datasets.

Conclusions:

• Federated learning is a viable, privacy-preserving approach for AI-based melanoma and nevus classification using distributed histopathological data.
• FL facilitates cross-institutional collaboration and can potentially be extended to other digital cancer histopathology tasks.
• While not outperforming ensemble methods, FL demonstrates potential for robust AI development in healthcare while safeguarding patient privacy.