ADGPT: A Prior-Guided and GPT-Empowered Framework for MRI-Based Alzheimer's Disease Diagnosis

Existing artificial intelligence methods for Alzheimer's disease (AD) diagnosis predominantly emphasize classification performance while neglecting interpretability. Leveraging large language models (LLMs) for diagnostic reasoning to generate decisions alongside supporting rationales can improve interpretability, yet remains underexplored in neuroimaging-based AD diagnosis. We propose a prior-guided and GPT-empowered framework (ADGPT) for magnetic resonance imaging (MRI)-based AD diagnosis. We construct an AD prior knowledge collection (APKC) to provide priors for the LLM's diagnostic reasoning. The APKC integrates three information types: label text tokens, JSON-based default mode network (DMN) representations, and brain tissue volumes. For the label text tokens, we propose a label text-graph structure alignment paradigm, such that the aligned label text tokens enable the LLM to identify AD-related abnormal functional patterns from functional MRI. The JSON-based DMN representations encode DMN functional connectivity in a structured JSON format, enabling the LLM to analyze the functional characteristics of key brain regions. Brain tissue volumes are used to quantify brain atrophy revealed by structural MRI. Additionally, we design a customized prompt for AD diagnosis to provide task inputs and guide the reasoning process. Without fine-tuning the LLM, ADGPT enables joint analysis of functional and structural AD pathologies revealed by neuroimaging and provides case-level, traceable diagnostic rationales. Experiments on two databases demonstrate that ADGPT outperforms state-of-the-art methods, achieving accuracies of 0.914 and 0.825 for AD versus normal control (NC), 0.806 and 0.812 for AD versus mild cognitive impairment (MCI), and 0.886 and 0.784 for MCI versus NC. Moreover, interpretability analyses provide additional evidence supporting its interpretability.