Abstract
Alzheimer's disease (AD) progression follows a complex continuum from normal cognition (NC) through mild cognitive impairment (MCI) to dementia, yet most deep learning approaches oversimplify this into discrete classification tasks. This study introduces M$^3$AD, a novel multi-task multi-gate mixture of experts framework that jointly addresses diagnostic classification and cognitive transition modeling using structural MRI. We incorporate three key innovations: (1) an open-source T1-weighted sMRI preprocessing pipeline, (2) a unified learning framework capturing NC-MCI-AD transition patterns with demographic priors (age, gender, brain volume) for improved generalization, and (3) a customized multi-gate mixture of experts architecture enabling effective multi-task learning with structural MRI alone. The framework employs specialized expert networks for diagnosis-specific pathological patterns while shared experts model common structural features across the cognitive continuum. A two-stage training protocol combines SimMIM pretraining with multi-task fine-tuning for joint optimization. Comprehensive evaluation across six datasets comprising 12,037 T1-weighted sMRI scans demonstrates superior performance: 95.13% accuracy for three-class NC-MCI-AD classification and 99.15% for binary NC-AD classification, representing improvements of 4.69% and 0.55% over state-of-the-art approaches. The multi-task formulation simultaneously achieves 97.76% accuracy in predicting cognitive transition. Our framework outperforms existing methods using fewer modalities and offers a clinically practical solution for early intervention.
Introduction
Alzheimer’s Disease (AD) is not a binary condition; it is a complex continuum moving from Normal Cognition (NC) through Mild Cognitive Impairment (MCI) to clinical Dementia. Most existing AI models oversimplify this by treating it as a discrete classification task, often employing "One-Versus-One" (OVO) strategies.
This "static" approach has two major flaws. First, it can lead to conflicting predictions, where models trained on different OVO tasks produce contradictory results for the same patient, lacking mutual exclusivity. Second, it ignores the transition, failing to capture the longitudinal "trajectories" of the disease. Understanding whether a patient is stable or likely to "convert" to a more severe stage is often more clinically valuable than a simple static diagnosis.
The M$^3$AD framework addresses these issues by treating AD as a continuous progression. By jointly learning diagnostic classification and cognitive transition patterns, the model ensures label consistency and provides a more comprehensive clinical picture.
Methodology
The M$^3$AD framework introduces several architectural and procedural innovations designed to extract fine-grained pathological features from structural MRI (sMRI).
Enhanced Swin Transformer V2 and Tok-MLP
The backbone of the model is an enhanced Swin Transformer V2. To improve sensitivity to regional brain atrophy while maintaining efficiency, the authors replaced the final stages of the transformer with Tok-MLP (Token-based Multi-Layer Perceptron) components. This modification enhances the model's ability to capture fine-grained structural changes while reducing the parameter count by nearly 50% compared to standard architectures.
Multi-gate Mixture of Experts (MMoE)
At the heart of the framework is the MMoE architecture. Instead of a single monolithic network, the model utilizes a set of "expert" networks. Some experts are shared to learn common structural features across the cognitive continuum, while others are specialized for specific diagnostic categories (e.g., NC-focused vs. AD-focused patterns). Dual "gates" dynamically route information to the most relevant experts for two simultaneous tasks: 3-class diagnosis (NC vs. MCI vs. AD) and transition pattern modeling.
Clinical Prior Integration and Two-Stage Training
The model integrates demographic data—age, gender, and estimated total intracranial volume (eTIV)—as clinical priors. These are fused with image features to improve generalization across diverse patient populations. The training protocol consists of two stages:
- SimMIM Pre-training: A self-supervised stage where the model learns to reconstruct masked brain scans. Label-guided reconstruction is used to force specific experts to specialize in certain pathological patterns.
- Multi-task Fine-tuning: The model is jointly optimized for both diagnostic classification and longitudinal transition modeling.
Experimental Results
The study utilized a massive dataset of 12,037 T1-weighted sMRI scans across six diverse cohorts, including ADNI, DLBS, IXI, NKI-RS, and OASIS-1/2. This large-scale evaluation ensures the model generalizes well across different scanners and acquisition protocols.
Diagnostic Performance
M$^3$AD set new benchmarks for sMRI-based analysis:
- 3-Class Diagnosis (NC/MCI/AD): Achieved 95.13% accuracy, a 4.69% improvement over the previous state-of-the-art (MCLNC).
- Binary Diagnosis (NC vs. AD): Achieved 99.15% accuracy.
- Transition Prediction: Simultaneously attained 97.76% accuracy in predicting whether a patient’s condition would remain stable or convert.
Modality Efficiency
The framework outperformed many existing methods that require more expensive or invasive modalities (such as PET scans or CSF biomarkers) while relying solely on standard sMRI and basic demographics. Ablation studies confirmed that removing demographic priors caused a significant drop in accuracy (~2%), highlighting that clinical context is a vital biomarker.
Conclusion
The success of M$^3$AD highlights the synergy between diagnostic and transition modeling. By treating AD as a continuum, the model learns more robust features that stabilize the diagnostic process. While the model shows exceptional performance, limitations remain regarding the scarcity of "reversion" data (patients moving from MCI back to NC). Future work will explore integrating genetic data into this MMoE framework to further refine early-stage detection. M$^3$AD provides a clinically practical, high-performance solution for AD diagnosis and prognosis, supported by an open-source preprocessing pipeline to address the field's reproducibility challenges.
Summary updated