Towards Explainable Multi-Modal AI for Alzheimer’s Diagnosis: A Comparative Review of Vision Transformers, Grad-CAM, and Language Models

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder in which early and reliable diagnosis remains one of the major clinical challenges. Recent advances in deep learning, particularly transformer-based architectures and large language models, have opened new opportunities for improved image-based diagnosis, explainability, and automated clinical reporting. However, the published work is scattered: most studies address one of the following aspects in isolation-image classification, visual explanation, multi-modal fusion, or report generation-with limited interlinking between the components. This paper presents a comparative review of recent research efforts on Vision Transformers (ViT) for MRI-based AD classification, explainable AI approaches such as Grad-CAM for neuroimaging interpretability, demographic and clinical data fusion, and transformer-based language models for radiology-style reporting. For each area, typical datasets are summarised, architectures, evaluation metrics, and reported strengths and further highlight key methodological and translational limitations related to data scarcity, lack of standardised explanation evaluation, multimodal integration challenges, and factual reliability issues in generated text. Based on this analysis, a future research direction towards an explainable is outlined, multimodal framework conceptually combining ViT, Grad-CAM, and domain-adapted language models for AD diagnosis and reporting without presenting new experimental results. We aim to map the current landscape, identify critical gaps, and motivate the design of more transparent and clinically aligned AI systems for Alzheimer’s disease. Keywords— Alzheimer’s Disease, Vision Transformer, Grad-CAM, Large Language Models, Explainable AI, MRI, literature review