STUDY SPACE
Federated Learning for Decentralized Artificial Intelligence in Melanoma Diagnostics

Federated Learning for Decentralized Artificial Intelligence in Melanoma Diagnostics

Published online February 7, 2024. | Sarah Haggenmüller, MSc; Max Schmitt, MSc; Eva Krieghoff-Henning, PhD; Achim Hekler, MSc; Roman C. Maron, MSc; Christoph Wies, MSc; Jochen S. Uttikal, MD; Friedegund Meier, MD; Sarah Hobelsberger, MD; Frank F. Gellrich, MD; Mildred Sergon, MD; Axel Hauschild, MD; Lars E. French, MD; Lucie Heinzinger, MD; Justin G. Schlager, MD; Kamran Ghoreshi, MD; Max Schlaak, MD; Franz J. Hilke, PhD; Gabriela Poch, MD; Sören Korsing, MD; Carola Berking, MD; Markus V. Heppt, MD; Michael Erdmann, MD; Sebastian Haferkamp, MD; Konstantin Drexler, MD; Dirk Schadendorf, MD; Wiebke Sondermann, MD; Matthias Goebeler, MD; Bastian Schilling, MD; Jakob N. Kather, MD; Stefan Fröhling, MD; Titus J. Brinker, MD
This study investigates the use of federated learning (FL) for melanoma-nevus classification, addressing privacy concerns associated with centralized AI models. The research involved 1025 whole-slide images from 923 patients, including 388 invasive melanomas and 637 nevi, collected from six German university hospitals. The primary endpoint was the area under the receiver operating characteristic curve (AUROC), with secondary endpoints including balanced accuracy, sensitivity, and specificity. The study compared the performance of FL with classical centralized learning and ensemble learning approaches. The results showed that the centralized approach outperformed FL on the holdout test dataset (AUROC: 0.9024 vs 0.8579), while FL performed better on the external test dataset (AUROC: 0.9126 vs 0.9045). The ensemble approach consistently outperformed both FL and the centralized approach on both datasets (AUROC: 0.9227 vs 0.9126 and 0.9045). The findings suggest that FL can achieve comparable diagnostic performance to classical centralized and ensemble approaches, making it a viable alternative for melanoma diagnostics. FL also enhances privacy protection and promotes collaboration across institutions and countries. The study highlights the potential of FL in other image classification tasks in digital cancer histopathology.This study investigates the use of federated learning (FL) for melanoma-nevus classification, addressing privacy concerns associated with centralized AI models. The research involved 1025 whole-slide images from 923 patients, including 388 invasive melanomas and 637 nevi, collected from six German university hospitals. The primary endpoint was the area under the receiver operating characteristic curve (AUROC), with secondary endpoints including balanced accuracy, sensitivity, and specificity. The study compared the performance of FL with classical centralized learning and ensemble learning approaches. The results showed that the centralized approach outperformed FL on the holdout test dataset (AUROC: 0.9024 vs 0.8579), while FL performed better on the external test dataset (AUROC: 0.9126 vs 0.9045). The ensemble approach consistently outperformed both FL and the centralized approach on both datasets (AUROC: 0.9227 vs 0.9126 and 0.9045). The findings suggest that FL can achieve comparable diagnostic performance to classical centralized and ensemble approaches, making it a viable alternative for melanoma diagnostics. FL also enhances privacy protection and promotes collaboration across institutions and countries. The study highlights the potential of FL in other image classification tasks in digital cancer histopathology.
Terms of Service
Privacy Policy
Reach us at info@study.space