Abstract
Neuroimaging studies based on magnetic resonance imaging (MRI) typicallyemploy rigorous forms of preprocessing. Images are spatially normalized to astandard template using linear and non-linear transformations. Thus, one canassume that a patch at location (x, y, height, width) contains the same brainregion across the entire data set. Most analyses applied on brain MRI usingconvolutional neural networks (CNNs) ignore this distinction from naturalimages. Here, we suggest a new layer type called patch individual filter (PIF)layer, which trains higher-level filters locally as we assume that moreabstract features are locally specific after spatial normalization. We evaluatePIF layers on three different tasks, namely sex classification as well aseither Alzheimer's disease (AD) or multiple sclerosis (MS) detection. Wedemonstrate that CNNs using PIF layers outperform their counterparts inseveral, especially low sample size settings.