A longitudinal and explainable 2.5D deep learning framework for Alzheimer's disease progression using ADNI MRI.

Early identification of Alzheimer's disease (AD) and its prodromal stage, mild cognitive impairment (MCI), is important for timely clinical assessment and disease management. Structural T1-weighted magnetic resonance imaging (MRI) captures macroscopic neurodegenerative changes associated with disease progression; however, developing deep learning models that are both methodologically rigorous and clinically interpretable remains challenging. This study presents an explainable deep learning framework for longitudinal classification of cognitively normal (CN), MCI, and AD subjects using MRI data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). A two-and-a-half-dimensional (2.5D) convolutional neural network based on a modified ResNet-18 architecture was used to incorporate contextual information from adjacent sagittal slices while remaining computationally efficient. To preserve longitudinal validity and reduce potential information leakage, multiple follow-up visits per subject were included and a strict subject-level data splitting strategy was adopted. Under this evaluation protocol, the proposed model achieved moderate classification performance, highlighting the difficulty of three-class longitudinal neurodegenerative disease classification under a stringent split design. To further examine the effect of split level, we additionally evaluated the same framework using a scan-level splitting strategy. In our implementation, scan-level splitting did not produce a substantial performance improvement, supporting the use of subject-level splitting primarily as a methodological choice for longitudinal validity. To enhance transparency, Gradient-weighted Class Activation Mapping (Grad-CAM) was used to interpret model predictions. The resulting explanations showed plausible temporo-frontal attention patterns in correctly classified AD cases and more heterogeneous responses in MCI cases. Overall, this work demonstrates that combining longitudinal MRI analysis with explainable deep learning under a rigorous evaluation framework can provide an interpretable and reproducible baseline for Alzheimer's disease research. By jointly emphasizing interpretability, temporal consistency, and evaluation validity, the proposed framework contributes toward the development of more trustworthy AI methods in neuroimaging, while also highlighting the remaining challenges in robust AD classification.