A Deep Learning-Based Approach for Early Alzheimer's Disease Diagnosis: Addressing Generalization and Domain Shift Across Populations

Abstract

Alzheimer’s Disease (AD) remains a leading cause of cognitive decline and dementia worldwide, affecting millions and placing a heavy burden on healthcare systems and caregivers. As the disease progresses over time, it often goes undiagnosed until moderate or late stages, where therapeutic interventions have diminished effectiveness. Timely diagnosis remains the single most impactful intervention, capable of significantly delaying disease progression. However, current diagnostic approaches largely identify AD only in its later stages, missing the critical opportunity for preventive action during the mild cognitive impairment (MCI) phase—a transitional stage where intervention is most beneficial. Therefore, advances in neuroimaging and artificial intelligence, particularly deep learning, have created potential opportunities for early detection of AD and prediction of MCI progression. This thesis explores deep learning as a promising approach in the early detection and progression prediction of Alzheimer’s Disease using structural magnetic resonance imaging (sMRI), addressing key limitations in interpretability, generalization, and data heterogeneity across medical cohorts. Chapter 1 begins by outlining the clinical and neuropathological basis of Alzheimer’s Disease, underscoring the importance of early detection through biomarkers such as hippocampal atrophy, amyloid-beta plaque accumulation, and tau protein tangles. It also introduces recent progress in neuroimaging-based diagnosis and the background of artificial intelligence, particularly deep learning. Structural Magnetic Resonance Imaging (MRI), with its high spatial resolution and soft tissue contrast, enables visualization of cerebral atrophy with significant potential for early diagnosis. Deep learning models, specifically convolutional neural networks (CNNs), have shown remarkable performance in disease classification tasks. This chapter sets the foundation for the subsequent work by building upon and expanding state-of-the-art techniques. Chapter 2 introduces the application of Vision Transformers (ViT) in modeling MCI-to-AD progression, as explored in the study “Vision Transformers for the Prediction of Mild Cognitive Impairment to Alzheimer’s Disease Progression using Mid-Sagittal sMRI.” Departing from CNN-based architectures, this work embraces ViTs for their ability to model global image dependencies using self-attention mechanisms. The model utilizes mid-sagittal slices from structural MRI, targeting brain regions such as the thalamus, medial frontal cortex, and occipital lobe—areas known to be pathologically relevant in MCI conversion. Using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), the proposed ViT framework achieves high accuracy (83.27%), specificity (85.07%), and sensitivity (81.48%) in distinguishing progressive MCI (pMCI) from stable MCI (sMCI). Importantly, attention maps derived from the transformer layers illuminate which brain regions most influence prediction, thereby enhancing model interpretability. This chapter illustrates the viability of attention-driven architectures in early AD diagnosis. Chapter 3 addresses a critical challenge in AD-related AI studies: data leakage due to improper splitting of training and test datasets. The study “Deep Learning with Guided Attention for Early Diagnosis of Alzheimer’s Disease” introduces the Guided-Attention Deep Learning (GADL) model that explicitly avoids this issue by enforcing subject-level data separation and integrates domain-specific attention with a CNN backbone. The GADL framework leverages a two-part architecture: a guided-attention feature extractor pretrained on the AD vs CN task, and a task-specific classifier for CN vs MCI and pMCI vs sMCI. The innovation lies in its guided-attention mechanism, which leverages knowledge distilled from a CN vs. AD classifier to identify the most informative regions for MCI classification. With subject-level splitting to prevent data leakage, the model achieved 80.29% accuracy for MCI progression prediction on the ADNI dataset and demonstrated high generalizability with 79.38% accuracy on the AIBL dataset. Chapter 4 expands upon the limitations of generalization posed by domain shift—variations in imaging protocols, demographics, and scanner settings that often degrade model performance when deployed across national cohorts. The third study, Domain Adaptation Network with Guided Attention for Multi- Cohort Alzheimer’s Disease Prediction proposes a semi-supervised domain adaptation technique that leverages guided attention scores from AD vs. CN classification to facilitate adaptation between source and target domains. The proposed architecture combines multi-plane sMRI feature extraction (sagittal, axial, and coronal) with semi-supervised domain adaptation and guided attention. Guided attention ensures that domain adaptation focuses on critical brain regions, while adversarial and pseudo-labeling strategies align feature distributions. The model is evaluated on both the ADNI and a separate national cohort with heterogeneous imaging settings, outperforming state-of-the-art baselines. The attention maps generated reveal consistent pathological focus areas across domains, underscoring the model’s robustness and reliability. Chapter 5 presents the most comprehensive and generalizable framework in this thesis: AttCORAL – Domain-Adaptive Attention Networks for Early Alzheimer’s Disease Diagnosis. Building upon prior efforts, this study incorporates Correlation Alignment (CORAL) loss with dual attention mechanisms in a residual neural network (ResNet) architecture. This approach combines the strengths of attention-based deep learning with statistical domain alignment through Correlation Alignment (CORAL) loss. The model architecture builds upon ResNet backbones enhanced with dual attention—spatial and channel— and integrates CORAL loss at the feature level to align the source and target domain distributions. Evaluated across both ADNI and AIBL datasets, AttCORAL achieves superior performance in both CN vs MCI classification and MCI-to-AD prediction tasks. What sets this model apart is its interpretability and adaptability: attention maps visualize the hippocampus, entorhinal cortex, and temporal lobe as key contributors to diagnosis, while CORAL ensures consistent performance across imaging settings and populations. The success of AttCORAL validates the thesis’ central argument—that attention-guided, domain-adaptive learning can lead to robust, interpretable, and generalizable AI models for neurodegenerative disease diagnosis. In summary, this thesis presents a cohesive, iterative exploration of deep learning methods tailored to the early diagnosis of Alzheimer’s Disease with a specific focus on MCI progression prediction. Each successive chapter builds upon the limitations of the previous, gradually improving accuracy, interpretability, and domain generalization. It demonstrates how attention mechanisms, domain adaptation, and rigorous validation protocols can collectively lead to high-performance, clinically meaningful, and ethically robust diagnostic models. As such, the work contributes significantly to the growing field of AI-driven neurodiagnostics and brings us closer to scalable, trustworthy tools for the early management of Alzheimer’s Disease.