Early diagnosis of Alzheimer's disease (AD), particularly at the mild cognitive impairment (MCI) stage, is crucial for time ly intervention. While volumetric neuroimaging modalities such as MRI, FDG PET, Tau PET, and Amyloid-B PET have demonstrated effectiveness in AD diagnosis, conventional 3D processing methods often fail to fully exploit their rich spatial information. Surface-based approaches, which provide more anatomically faithful representations, have shown promise but typically rely on single modalities and suffer from costly inference due to complex preprocessing pipelines.

This thesis first presents a unified multimodal framework that bridges the advantages of both volume and surface representations. By employing knowledge distillation, the proposed method utilizes surface data only during training to enhance volumetric inference, enabling fast and cost-effective deployment. A novel architecture incorporating multi-view and multi-convolution blocks is introduced to extract rich features from diverse modalities. Comprehensive evaluations on the ADNI dataset series confirm the robustness and generalization capability of the approach, achieving state-of-the-art results across various PET and MRI modalities.

Building on findings from this study, where cortical surfaces were shown to provide valuable information and demonstrated advantages in multiple aspects, I conducted a second study focusing solely on surface-based modeling. Specifically, I proposed ASAM, an adaptive cortical surface modeling framework that integrates local spatial awareness into Mamba-based sequence modeling. To overcome the limitations of fixed receptive fields and traversal-dependent representations on non-Euclidean cortical meshes, ASAM introduces Dynamic Multi-Hop Selection and Dynamic Feature Merging modules. These components adaptively expand neighborhood aggregation based on geodesic and feature affinities, enabling precise characterization of complex cortical morphologies. Experiments on standard brain surface benchmarks demonstrate that ASAM significantly outperforms leading Transformer- and Mamba-based baselines, offering a scalable and anatomically grounded solution for surface-based neuroimaging tasks.