Data-assisted differential diagnosis of dementia by deep neural networks

Abstract

There are currently 50 million people suffering from dementia worldwide. With an increasing life expectancy of the elderly, this number is expected to increase drastically over the next decade. With today's diagnosis of dementia being highly dependent on the expertise of clinical personnel, there is thus a pressing need for readily available and reliable diagnosis systems.

In this thesis, the potential for computer-aided diagnosis systems based on deep neural networks and structural magnetic resonance imaging is investigated. An ensemble model was designed and trained on 690 class-balanced brain scans for the differentiation of subjects diagnosed with Alzheimer's disease and dementia with Lewy bodies, as well as normal control subjects. All scans were initially skull-stripped and spatially normalized to remove unwanted information.

A final accuracy of 71.9% was reported for the three class differentiation of 171 test subjects. The results presented in this thesis fall a little short in comparison with those of related work, but indicates, nonetheless, a potential for this type of deep learning-based diagnosis systems.