Bayesian XAI Methods Towards a Robustness-Centric Approach to Deep Learning: An ABIDE i Study

Developing reliable and explainable models is a crucial point to effectively integrate and exploit the potentials offered by Deep Learning (DL) architectures in high-stakes scenarios like healthcare. There are several applications that exploit DL to support Autism Spectrum Disorder (ASD) diagnosis, eventually augmented by explainable AI (XAI) tools to provide hints on the decision-making process implemented. On the other hand, Bayesian Neural Networks (BNNs) can provide, together with their prediction, epistemic uncertainty (uncertainty of the model), a key component to asserting the model's reliability. To date, there are no applications which exploit the advantages offered by BNNs and XAI to support the research of biomarkers in ASD. In the present work, authors first developed a BNN which classifies ASD subjects from resting state functional Magnetic Resonance Imaging (rs-fMRI) data obtained from the Autism Brain Imaging Data Exchange (ABIDE) dataset. A Layerwise Relevance Propagation (LRP) algorithm was then used to estimate the importance of cross-correlation connectivity coefficients in the returned predictions. Finally, a group analysis was performed to highlight functional brain connections that report the highest impact on the model's correct classification of ASD subjects. This work ended up producing a framework which combines a bayesian neural network with a XAI methodology, towards a robustness-centric deep learning approach, applied to the case study of ASD diagnosis.