Joint functional brain network atlas estimation and feature selection for neurological disorder diagnosis with application to autism

Image-based brain maps, generally coined as ‘intensity or image atlases’, have led the field of brain mapping in health and disease for decades, while investigating a wide spectrum of neurological disorders. Estimating representative brain atlases constitute a fundamental step in several MRI-based neurological disorder mapping, diagnosis, and prognosis. However, these are strikingly lacking in the field of brain connectomics, where connectional brain atlases derived from functional MRI (fRMI) or diffusion MRI (dMRI) are almost absent. On the other hand, conventional connectomic-based classification methods traditionally resort to feature selection methods to decrease the high-dimensionality of connectomic data for learning how to diagnose new patients. However, these are generally limited by high computational cost and a large variability in performance across different datasets, which might hinder the identification of reproducible biomarkers. To address both limitations, we unprecedentedly propose a brain network atlas-guided feature selection (NAG-FS) method to disentangle the healthy from the disordered connectome. To this aim, given a population of brain connectomes, we propose to learn how estimate a centered and representative functional brain network atlas (i.e., a population center) to reliably map the functional connectome and its variability across training individuals, thereby capturing their shared traits (i.e., connectional fingerprint of a population). Essentially, we first learn the pairwise similarities between connectomes in the population to map them into different subspaces. Next, we non-linearly diffuse and fuse connectomes living in each subspace, respectively. By integrating the produced subspace-specific network atlases we ultimately estimate the population network atlas. Last, we compute the difference between healthy and disordered network atlases to identify the most discriminative features, which are then used to train a predictive learner. Our method boosted the classification performance by 6% in comparison to state-of-the-art FS methods when classifying autistic and healthy subjects.