DeMed: A Novel and Efficient Decentralized Learning Framework for Medical Images Classification on Blockchain

Garima Aggarwal, Chun-Yin Huang, Di Fan, Xiaoxiao Li, Zehua Wang

Abstract:

Training predictive models with decentralized medical data
can boost the healthcare research and is important for healthcare applications.
Although the federated learning (FL) was proposed to build
the predictive models, how to improve the security and robustness of a
learning system to resist the accidental or malicious modification of data
records are still the open questions. In this paper, we describe DeMed, a
privacy-preserving decentralized medical image analysis framework empowered
by blockchain technology. While blockchain is limited in serial
computing, the decentralized data interaction in blockchain is very
desired to preserve the data privacy when training models. To adapt
blockchain in medical image analysis, our framework consists of the self-supervised
learning part running on users’ local devices and the smart
contract part running on blockchain. The prior is to obtain the provable
linearly separable low-dimensional representations of local medical images
and the latter is to obtain the classifier by synthetically absorbing
users’ self-supervised learning results. The proposed DeMed is validated
on two independent medical image classification tasks on pathological
data and chest X-ray. Our work provides an open platform and arena for
FL, where everyone can deploy a smart contract to attract contributors
for medical image classification in a secure and decentralized manner
while preserving the privacy in medical images.