NIMG-68. FEDERATED LEARNING IN NEURO-ONCOLOGY FOR MULTI-INSTITUTIONAL COLLABORATIONS WITHOUT SHARING PATIENT DATA. (11th November 2019)

• Record Type: Journal Article
• Authors: Sheller, Micah; Edwards, Brandon; Anthony Reina, G; Martin, Jason; Bakas, Spyridon
• Is Part Of: Neuro-oncology. Volume 21(2019)Supplement 6
• Page Start: vi176
• Page End: vi177
• Publication Date: 2019-11-11
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noz175.737
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: BACKGROUND: Training deep learning algorithms requires large amounts of data, which is a significant challenge in the medical domain, and particularly in neuro-oncology, where ample data can only be found in multi-institutional collaborations. The current paradigm for multi-institutional collaborations is based on pooled datasets that has always faced privacy, legal, technical, and data-ownership concerns. In this study we evaluate the hypothesis that federated learning can provide a method to overcome these concerns and facilitate a shift in the paradigm of multi-institutional collaborations without sharing patient data. We attempt to investigate this hypothesis in a feasibility study of automatically delineating the glioblastoma extent in T2-FLAIR scans. METHODS: We identified a retrospective cohort of 165 glioblastoma patients with available clinically acquired pre-operative multi-parametric structural MRI (mpMRI) scans (i.e., T1, T1Gd, T2, T2-FLAIR), with corresponding expert tumor boundary annotations, from 10 independent institutions. We implemented a 3D deep learning algorithm (3D-UNet) to predict the boundaries of the whole tumor extent, by virtue of the abnormal hyper-intense signal of T2-FLAIR scans. We compare the performance of this 3D-UNet model resulting from federated learning with the performance of the same 3D-UNet model generated by sharing data to a single location where centralized/traditional training occurs. RESULTS: Our quantitative results on federated learning (Dice:85.2%) across individual contributions from the 10 institutions demonstrate final model quality reaching 99% of the model quality achieved by sharing data (Dice:86.2%). CONCLUSIONS: Translation and adoption of federated learning in a clinical configuration for multi-institutional collaborations is expected to have a catalytic impact towards precision and personalized medicine. The performance of computer-aided analytics and assistive diagnostics is expected to see a precipitous rise, as new models are trained on datasets of unprecedented size through such data-private collaborations.