Integrated Gradients Demystified: An MRI Case Study on Aβ-T Protein Localization

Explainability in Artificial Intelligence is gaining increasing importance, especially in critical fields. In biomedical applications, attribution maps are particularly relevant for their inherent spatial localization property. Features that are relevant for a given task (e.g., classification of cases versus controls) can then be associated to specific regions, revealing disease-specific modulations and enabling ad-hoc post-hoc validation. Among these, baseline-dependent methods require a neutral image (called baseline) for computing the feature importance. The choice of the baseline is critical in that it determines the relevance map. In this work, we focused on a classification task, that is, subject stratification based on the amyloid-β and τ status in controls and patients in the Alzheimer's disease continuum, and employed Integrated Gradients to extract the attribution maps. This allowed identifying relevant brain regions driving the classification. Aware of the dependence of the outcomes from the chosen baseline, post-hoc statistical analysis was performed for validating the results confirming their plausibility. This puts forth the need to systematically assess the impact of the baseline for the correct interpretation of the corresponding outcomes.