Federated Split Vision Transformer for COVID-19 CXR Diagnosis using\n Task-Agnostic Training

Abstract

Federated learning, which shares the weights of the neural network across\nclients, is gaining attention in the healthcare sector as it enables training\non a large corpus of decentralized data while maintaining data privacy. For\nexample, this enables neural network training for COVID-19 diagnosis on chest\nX-ray (CXR) images without collecting patient CXR data across multiple\nhospitals. Unfortunately, the exchange of the weights quickly consumes the\nnetwork bandwidth if highly expressive network architecture is employed.\nSo-called split learning partially solves this problem by dividing a neural\nnetwork into a client and a server part, so that the client part of the network\ntakes up less extensive computation resources and bandwidth. However, it is not\nclear how to find the optimal split without sacrificing the overall network\nperformance. To amalgamate these methods and thereby maximize their distinct\nstrengths, here we show that the Vision Transformer, a recently developed deep\nlearning architecture with straightforward decomposable configuration, is\nideally suitable for split learning without sacrificing performance. Even under\nthe non-independent and identically distributed data distribution which\nemulates a real collaboration between hospitals using CXR datasets from\nmultiple sources, the proposed framework was able to attain performance\ncomparable to data-centralized training. In addition, the proposed framework\nalong with heterogeneous multi-task clients also improves individual task\nperformances including the diagnosis of COVID-19, eliminating the need for\nsharing large weights with innumerable parameters. Our results affirm the\nsuitability of Transformer for collaborative learning in medical imaging and\npave the way forward for future real-world implementations.\n

Ähnliche Arbeiten