Review of deep learning models for Alzheimer's disease detection: MRI-centric approaches and multimodal extensions.

Introduction: Alzheimer's disease (AD) is a leading cause of dementia worldwide, and an early, reliable diagnosis is critical for timely intervention. Structural magnetic resonance imaging (MRI), coupled with deep learning (DL), has emerged as a promising non-invasive approach for automated diagnosis. This review evaluates DL models applied to MRI for AD detection, while also considering multimodal extensions (PET, fMRI, DTI, CSF, and cognitive data) that augment MRI-based pipelines. Methods: Following PRISMA guidelines, a comprehensive search of six databases (2010-June 2025) identified 70 peer-reviewed studies, with many of them integrating multimodals as well. Data on model architectures, datasets, pre-processing, validation protocols, and reported performance outcomes were extracted and synthesised. Results: Most studies have employed 2D or 3D convolutional neural networks; however, recent work has also explored ensembles, vision transformers, graph neural networks, and generative models. ADNI was the primary dataset, in addition to OASIS, AIBL, UK Biobank, and other cohorts that were also utilised. Binary classification tasks distinguishing clinically diagnosed Alzheimer's disease (AD) patients from cognitively normal (CN) controls consistently reported high performance (>90% accuracy, AUC ≥ 0.95). In contrast, more clinically challenging tasks, such as multiclass classification across disease stages (CN, MCI, AD) and prediction of MCI-to-AD conversion, yielded substantially lower accuracy (approximately 70-85%). Reported near-perfect results (>99%) were often confined to single-site datasets lacking external validation, raising concerns of overfitting and reproducibility. Discussion: Few studies have incorporated differential diagnosis of dementia, advanced harmonisation across scanners, or open-source pipelines. Promising advances include transfer learning, multimodal integration, harmonisation methods (such as ComBat, GANs, diffusion), and explainable AI techniques. Overall, DL shows strong potential for MRI-based AD detection, with multimodal inputs further improving performance.