Distributed learning for optimal radiomics knowledge

This thesis evaluated the current state of research in the field of radiomics and presented an up-to-date overview of distributed learning applications in health care and their limitations. It also presented a new infrastructure to address the limitations of the existing distributed learning solution. The potential of decentralizing the learning process by enforcing trust, immutability, transparency, and traceability in distributed learning networks via blockchain smart contracts was studied. The results showed that the decentralization of distributed learning networks has the potential to train the models on patient data stored at different health care centres without need to exchange patient data or to trust a third party. Furthermore, it evaluated the potential of distributed learning on small sets of data that can mimic the case of low prevalence diseases and phase I clinical trials, where it is not possible for a single health care provider to hold enough data to train reliable AI models. The results suggest that distributed learning 1) overcomes data sharing limitations for low prevalence diseases and can be the new way to apply AI for such diseases, 2) is capable to train models as good as centrally trained models. Moreover, an infrastructure was deployed to a public blockchain (Ethereum) test network and tested the feasibility of decentralized distributed learning in real world settings. The decentralized nature of the study makes it easy for the public to evaluate the time, cost, and the learning quality, hence reaching transparent learning process.