Federated Learning Comparable to Traditional Learning Methods for AI-based Melanoma Diagnosis

The learning approach has the potential to extend to other classification tasks in digital cancer histopathology.

A decentralized federated learning approach was comparable in diagnostic performance to traditional centralized and ensemble learning methods in AI-based melanoma diagnosis, new research suggests.

Study authors conducted a multicentric, single-arm diagnostic study conducted across six German university hospitals between April 2021 and February 2023. The study used 1,025 histopathological whole-slide images of melanoma-suspicious skin lesions from 923 patients, including 388 confirmed invasive melanomas and 637 nevi. The primary endpoint for evaluation was the area under the receiver operating characteristic curve (AUROC), with secondary endpoints including balanced accuracy, sensitivity, and specificity.

The federated approach performed slightly worse than the centralized approach on a holdout test dataset, but significantly better on an external test dataset. The federated approach achieved an AUROC of 0.8579 (95% CI, 0.7693-0.9299) vs. 0.9024 (95% CI, 0.8379-0.9565) for the centralized approach on the holdout test dataset. For the external test dataset, the federated approach outperformed the centralized approach with an AUROC of 0.9126 (95% CI, 0.8810-0.9412) vs. 0.9045 (95% CI, 0.8701-0.9331). Notably, the federated approach performed worse than the ensemble approach on both test datasets.

"The results of this diagnostic study demonstrate that federated can achieve a comparable performance to that of classical centralized or ensemble approaches, making it a reliable alternative for the classification of IMs and nevi," the authors wrote. "Additionally, federated learning empowers institutions to contribute to the development of AI models, even with relatively small datasets or strict data protection rules, thereby fostering collaboration across institutions and countries."

The authors added that federated learning has the potential to be extended to other image classification tasks in digital cancer histopathology.

"Future research could build on this work by assessing its effectiveness with different types of medical images (eg, dermoscopic or hyperspectral images), evaluating its feasibility for diagnosing various types of cancer, and investigating its effectiveness using technically different (eg, attention-based methods) AI models," they wrote.

The study was published online in JAMA Dermatology.

Source: Haggenmüller S, Schmitt M, Krieghoff-Henning E, et al. Federated Learning for Decentralized Artificial Intelligence in Melanoma Diagnostics. JAMA Dermatology. Published online February 7, 2024. doi:https://doi.org/10.1001/jamadermatol.2023.5550