Modeling inter-slice dependencies with temporal graph learning for Alzheimer's disease.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Accurate early diagnosis of Alzheimer's disease (AD) is crucial but challenging, especially differentiating mild cognitive impairment (MCI) from cognitively normal (CN) aging.
• Conventional 2D models struggle to capture 3D MRI spatial context and inter-slice temporal dependencies.

Purpose of the Study:

• To propose TGL-AD, a novel framework integrating Vision Transformer (ViT) slice embeddings with temporal graph learning for subject-level AD classification.
• To enhance the accuracy of distinguishing AD, MCI, and CN using 3D MRI data.

Main Methods:

• 3D MRI volumes were decomposed into 2D slices and encoded using a pre-trained ViT to generate feature embeddings.
• Temporal graphs were constructed to model inter-slice continuity, with graph neural networks (GNNs) transmitting contextual information.
• Subject-level representations were generated for classification via global pooling.

Main Results:

• TGL-AD achieved high accuracy: 0.92 on the 1.5T cohort and 0.98 on the 3T cohort of the ADNI1 dataset.
• Consistently high precision, recall, and F1-scores were observed across AD, MCI, and CN diagnostic categories for both cohorts.
• TGL-AD outperformed leading CNN-based and transformer-based baselines, achieving superior recall and F1-scores.

Conclusions:

• The integration of ViT slice encoders and temporal graph modeling effectively captures inter-slice dependencies in MRI data.
• TGL-AD demonstrates enhanced classification performance for Alzheimer's disease diagnosis across different MRI acquisition settings.
• This framework offers a promising approach for improving early and accurate detection of neurodegenerative conditions like AD.