Constrained Source-based Salience: Network-based Visualization of Deep Learning Neuroimaging Models

Developing frameworks using high-dimensional magnetic resonance imaging (MRI) data to characterize underlying brain changes in neurological disorders is crucial and challenging. While deep learning models offer a better prediction, tracking automated higher-order explanations at the level of brain networks is harder in learned models. We introduce a novel constrained source-based salience (cSBS) framework to automatically learn and visualize multiple independently salient brain networks associated with clinical diagnostic assessments. This is achieved by performing active subspace learning (ASL) and spatially constrained independent component analysis (scICA) in the saliency space of trained convolutional neural networks (CNNs), such that the resultant components are interpretable in terms of brain network components from existing templates. By employing a robust analysis across repeated training scenarios for an Alzheimer's disease (AD) classification task, we visualize cSBS components via full-brain back-reconstruction. We show that the cSBS components and their corresponding loadings are consistent and relevant in terms of AD-related brain areas. Our approach is able to synthesize multiple objectives of utilization of high-dimensional MRI data for deep learning along with automated detection of low-dimensional representations of the consistently involved features in terms of intrinsically salient brain networks. Our framework of automated identification of consistent underlying brain subsystems associated with clinically observed assessments is an important step toward biomarker development for various clinically observed characteristics and disorders.