Photo Credit: Malikov Aleksandr

This diagnostic study investigates the performance of a privacy-preserving federated learning approach vs a classical centralized and ensemble learning approach.

The diagnostic performance of federated learning may be comparable to traditional centralized approaches for AI-based melanoma diagnostics, according to data recently published in JAMA Dermatology. In a comprehensive multicenter study analyzing 1,025 whole-slide images from 923 patients with clinically suspicious melanoma lesions, Sarah Haggenmüller, MSc, and colleagues, explored the efficacy of a privacy-centric federated learning model against traditional centralized and ensemble learning methods for AI-based melanoma diagnostics. The study showed that the federated approach performed significantly worse than the classical centralized approach in terms of AUROC on a holdout test dataset but performed significantly better than the classical centralized approach on an external test dataset. Although the federated approach performed significantly worse than the ensemble approach on both the holdout and external test dataset, the researchers suggest that federated learning is a promising method for classifying melanomas and nevi, enhancing privacy and collaboration in AI-based diagnostics, with potential applications in digital cancer histopathology and beyond.