Selection, visualization, and explanation of deep features from resting-state fMRI for Alzheimer's disease diagnosis

Despite the remarkable achievements of deep learning networks in analyzing neuroimaging data for various tasks linked to brain functions and disorders, the opaque nature of these models and their interpretability challenges pose significant barriers to their broader use in clinical settings. This research scrutinizes the visualization of deep features from resting-state functional magnetic resonance imaging (rs-fMRI) images to discriminate individuals who are cognitively normal from those with different stages of Alzheimer's disease. Rs-fMRI data are obtained from the ADNI database. This research indicates the presence of a specific subset of deep features capable of effectively identifying Alzheimer's, termed "informative deep features." By visualizing the distinct deep features, we gain better insights into the pathological patterns present at each level of the entire rs-fMRI volume, despite the challenges posed by closely resembling patterns of brain atrophy and image intensities. These deep features were visualized across the whole slide image level using deep feature-specific heatmaps and activation maps. Furthermore, the findings imply that these significant deep features may hold diagnostic potential for other central nervous system disorders beyond Alzheimer's. This framework could act as a basis for assessing the interpretability of any deep learning model in the context of diagnostic decision-making.