Multi-plane multi-slice longitudinal MRI for deep ensemble progression detection based on enhanced residual multi-head self-attention

Early and accurate detection of Alzheimer's disease (AD) progression remains a critical challenge in neuroimaging, as existing methods often rely on single-plane or cross-sectional MRI data, neglecting the rich spatiotemporal dynamics captured in multi-plane longitudinal imaging. To address this gap, we propose a novel deep ensemble framework that leverages multi-plane volumetric representations of 3D longitudinal MRI data for enhanced AD progression detection. Our approach introduces a unified 3D volumetric representation of longitudinal MRI by integrating spatially aligned slices from axial, coronal, and sagittal planes across four longitudinal time points (Baseline, M06, M12, and M18), preserving both spatial and temporal context. Our proposed framework employs an optimized heterogeneous deep ensemble setup of 3D-CNN models (i.e., 3D-EfficientNet, 3D-DenseNet, and 3D-ResNet) to extract complementary spatio-temporal features from each anatomical plane, followed by a BiLSTM network with an Enhanced Residual Multi-Head Self-Attention (ERMHA) mechanism to model long-range dependencies and emphasize discriminative spatiotemporal patterns. Comprehensive experiments on the ADNI dataset demonstrate that our proposed framework achieves state-of-the-art performance, with a mean accuracy of 93.73%, sensitivity of 91.72%, specificity of 90.36%, and an AUC of 91.58%, significantly outperforming single-plane based models (best mAUC: 68.24%) and homogeneous ensemble approaches (mAUC: 82.75%). External validation on the NACC cohort further confirms generalizability, with performance metrics improving consistently as data from more longitudinal time steps are incorporated (mAUC: 86.37% at M18). Furthermore, explainability analysis using gradient-weighted attention maps reveals that model predictions are driven by neuroanatomically plausible patterns, with attention focused on hippocampal and entorhinal regions in early progression and extending to temporo-parietal cortices in advanced stages in AD, aligning with established neuropathological trajectories. The proposed framework advances intelligent decision support systems in clinical neuroimaging by combining multi-plane feature fusion, temporal modeling, and ensemble learning, offering a robust and generalized solution for early AD progression detection. Its modular design and computational efficiency make it suitable for integration into knowledge-based diagnostic systems. The dataset and code used in this study are available to the research community at the following link: https://github.com/InfoLab-SKKU/mpms-mri-progression.git