AXIAL: Attention-based eXplainability for Interpretable Alzheimer’s Localized diagnosis using 2D CNNs on 3D MRI brain scans

Deep learning models for Alzheimer’s disease (AD) diagnosis from magnetic resonance imaging (MRI) have achieved high performance, but clinical adoption remains limited because reliability and generalizability remain uncertain. Reported results are often difficult to compare because cohorts and evaluation protocols differ, while model explanations are rarely assessed quantitatively. This study investigates whether attention-based models evaluated under a rigorous and reproducible protocol can provide reproducible performance estimates and biomarker-oriented population-level explanations. We propose AXIAL, an attention-based framework that classifies three-dimensional structural MRI using pretrained two-dimensional convolutional neural networks and learns slice importance to synthesize a voxel-level attention map from orthogonal slice-attention weights. Evaluation used the standardized ADNI1 Complete 1 Yr 1.5T collection with participant-separated five-fold cross-validation, participant-bootstrap confidence intervals, and participant-level sensitivity analyses. Generalizability was assessed through a locked retrospective evaluation in the AIBL cohort without adaptation. Explainability analyses included cross-fold stability, controlled slice deletion, and threshold sensitivity. For AD versus cognitively normal classification, AXIAL achieved an acquisition-level area under the receiver operating characteristic curve of 0.915 (95% confidence interval 0.887–0.940), accuracy of 0.857 (0.824–0.886), and Matthews correlation coefficient of 0.710 (0.644–0.771). In the locked external evaluation on AIBL, the area under the curve was 0.894 (0.844–0.940), with AD sensitivity of 0.783 (0.681–0.870) and specificity of 0.858 (0.826–0.890). For progressive versus stable mild cognitive impairment, with progressive cases treated as positive, the area under the curve was 0.719 (0.651–0.789), accuracy was 0.725 (0.671–0.782), and Matthews correlation coefficient was 0.449 (0.342–0.565). Attention maps emphasized the hippocampus, parahippocampus, and amygdala. Stability varied by anatomical plane; masking the highest-attention slices caused the greatest performance deterioration, and hippocampal overlap persisted across thresholds. Under transparent and reproducible evaluation, AXIAL provided reproducible predictive estimates and biomarker-aligned population-level attention outputs. These findings support the view that improving reliability of both predictive evaluation and model explanations, rather than maximizing isolated performance claims, is a key step toward clinically trustworthy automated AD diagnosis.