Federated Learning for Digital Pathology: A Pilot Study
Geetu Mol Babu, Kok Wai Wong, Jeremy Parry · Procedia Computer Science · 2022
Over the last few years, there have been many significant advances in the use of deep learning in digital pathology. Deep learning has been reported to assist with registration, segmentation, classification, and diagnosis tasks in digital pathology. However, the development of scalable, adaptable, and accurate deep learning-based models often relies on collecting large amounts of high-quality annotated training data from various sources or from different sites. Given that the medical data are normally linked to a site and/or within an organization, assembling large-scale datasets has historically required data transfer between them. Patient privacy could be an issue by such transfers, especially if data are to be shared between countries. This poses ethical and legal issues. Therefore, privacy and ownership issues affecting multi-institutional collaborations focused on centrally shared patient data, could affect the translation of deep learning techniques for real-world application. Federated learning has recently emerged as a new paradigm for data-private multi-institutional collaborations, in which model-learning uses all available data without exchanging data between institutions by distributing model training to data-owners and aggregating their outcomes. This paper aims to investigate the possible use of the federated learning method in digital pathology and examine the advantages of using federated learning for real world digital pathology workflow. We have performed our study on the Breast Cancer Histopathological Database, which consists of data from different sites. The case study results presented in this paper have demonstrated that federated learning can be used effectively in the digital pathology area.