Federated Learning for Dementia Classification in a European Multicentre Dementia Study

Abstract

Every year around 10 million people are diagnosed with dementia worldwide. Higher life expectancy and population growth could inflate this number even further in the near future. Currently the diagnostic process of dementia relies heavily on medical experts on an individual basis. As the prevalence of the disease grows, so does the need for reliable diagnosis systems. Medical institutions around the world hold massive amounts of medical patient data. Large portions of this data can not be shared between institutions due to patient privacy concerns. This thesis explores some solutions to these obstacles. Computer-aided diagnosis systems based on various deep neural networks trained on magnetic resonance imaging is investigated. The use of generative adversarial networks to generate usable samples for deep neural networks without compromising patient privacy is explored. A federated structuring of deep neural networks where patient data is kept locally is tested. Data for all experiments are based on a class-balanced dataset of 690 brain scans from patients diagnosed with Alzheimer’s disease, dementia with Lewy bodies and normal control subjects. An accuracy of 78.65% was achieved for a three class differentiation of 171 test subjects. This is a formidable result, especially compared to related deep learning based approaches. The generative adversarial network approach of generating new data achieved fairly good results, but due to memory limitations this data is of lower resolution and could not be used in the final evaluation. The federated structuring of deep neural networks yielded in part promising results and could be an important way of accessing medical data while protecting privacy in the future.