University of Twente Student Theses

Login

Split Learning in Health Care: Multi-center Deep Learning without sharing patient data

Poirot, Maarten G. (2020) Split Learning in Health Care: Multi-center Deep Learning without sharing patient data.

Full text not available from this repository.

Full Text Status:Access to this publication is restricted
Embargo date:31 December 2020
Abstract:Purpose: For many diseases, the sample sizes at any one hospital are often small, patient information cannot be shared openly, and multi-center collaborative studies are a burden to set up. Distributed machine learning methods promise to mitigate these problems. We investigate feasibility and opportunities of a novel distributed deep learning paradigm called Split Learning that partitions a conventional neural network in sequential elements that can be either private or centralized. Further research is presented on 1) handling of heterogeneous data by making use of Adapter Networks 2) improving privacy using alternative model sharing strategies and 3) how to handle vertically partitioned data. Methods: We implemented Split Learning in a simulated multi-institutional setting for several medical imaging tasks using the BraTS, CheXpert and MURA data sets, and compared the performance of Split Learning with that of centrally hosted data. We investigated the relationship between inference performance and convergence rate, and the number of participating institutions. Lastly, we compared Institution-side computational requirements and communicational cost of Split Learning compared to the most popular alternative method, namely, Federated Learning. Results: Adaptation to Split Learning was successful for all tasks. Inference performance and convergence rate achieved using Split Learning were not affected by an increased number of participating institutions. Institution-side computational resource requirements were significantly lowered using Split Learning. Communication bandwidth requirements for Split Learning, although higher than Federated Learning, were well within the data communication rates offered by modern internet connectivity. Conclusions: We show that Split Learning can be used to train deep learning models for medical imaging without the need to share patient data. We also demonstrate that Split Learning can reduce the local computational resource requirements at each institution.
Item Type:Essay (Master)
Clients:
Massachusetts General Hospital, Boston, MA, United States of America
Massachusetts Institute of Technology, Cambridge, MA, United States of America
Faculty:TNW: Science and Technology
Subject:54 computer science
Programme:Technical Medicine MSc (60033)
Link to this item:http://purl.utwente.nl/essays/80601
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page