Abstract
Abstract
AI-assisted early diagnosis of Alzheimer’s disease (AD) has substantial clinical value, and diffusion tensor imaging (DTI), which captures white matter microstructural alterations, has considerable potential across the AD continuum. However, major barriers to clinical translation remain. This review systematically evaluated the evolution of algorithmic paradigms, the effectiveness of multimodal fusion, and barriers to clinical generalization in DTI-based AD diagnosis. Using an “Input-Model-Fusion” evaluation framework, we analyzed 98 studies published from 2010 to 2026 and compared sample size, diagnostic task, validation design, and methodological quality. Four findings emerged. 1) Modality selection should be task-driven; single-modality DTI and multimodal fusion have different applicability across diagnostic tasks. 2) Algorithm choice was closely related to sample size: conventional machine learning, particularly support vector machines, predominated in small-sample studies, whereas deep learning architectures, including convolutional neural networks, graph convolutional networks, and Transformers, were more widely used with large datasets. 3) Substantial performance degradation in multicenter or external validation highlighted persistent limitations in clinical generalizability. Data source, sample representativeness, feature processing, validation strategy, acquisition protocol, and software pipeline heterogeneity remain important constraints on clinical translation and cross-study comparability of DTI metrics. 4) Hippocampal pathways, the corpus callosum, fornix, and cingulum showed relatively stable discriminative value. Future research should move beyond “algorithm accuracy competition” toward clinical pragmatism by prioritizing task alignment, standardized DTI acquisition and processing, external validation, and neurobiological interpretability.