Deep fusion pipeline for mild cognitive impairment diagnosis

Abstract

Deep learning has allowed scientists to make significant improvements in tasks that were once considered difficult in disparate domains. Medical imaging is one of those domains where traditional analysis entailed multiple preprocessing steps and feature extraction or handcrafting of individual features for specific applications. Deep learning allows one to simplify this analysis pipeline into an end-to-end framework as it can handle the feature extraction phase without having to handcraft features. We leverage this characteristic of deep learning and present an architecture where multiple information modalities of different complexities can be fused together seamlessly and co-optimized to create a robust classifier. The performance of this fusion pipeline is demonstrated on Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset where discrimination between Alzheimer's Disease, Mild Cognitive Impairment and cognitively normal individuals using 3D magnetic resonance imaging and neuropsychological measures is presented.