Review of Explainable AI Approaches in Multimodal Neuroimaging for Alzheimer’s Disease Detection

Alzheimer's disease is the most prevalent etiology of dementia, and its early diagnosis has been deemed instrumental for early intervention and better patient management. Over the last few years, deep learning techniques have been used with neuroimaging data, and notable gains in diagnostic performance have ensued. Such models have, however, typically been faulted for being "black boxes" that, despite their performance, hinder their acceptance in clinical practice due to a lack of transparency and interpretability. Current explainability methods have typically been used as post-hoc procedures, but they often yield inconsistent or anatomically irrelevant attribution maps that clinicians find hard to trust. In this review, progress in explainable AI for the detection of Alzheimer's is discussed, with a focus on DL methods combined with multimodal neuroimaging. Emphasis is given to XRAI, a region-based attribution method that has been demonstrated to produce more coherent and clinically interpretable explanations compared to conventional pixel-level techniques. The application of XRAI to 2D and 3D neuroimaging is considered, together with the potential of XRAI to identify anatomically relevant brain areas implicated in disease pathology. Challenges in terms of clinical uptake, integration into workflow, and standardized evaluation of explainability techniques are also discussed. The review identifies how the pairing of high-performing AI models with strong explainability methods has the potential to enable the creation of practical and reliable diagnostic tools for real-world clinical application.