ASAM: Adaptive cortical surface representation via dynamic local aggregation and hybrid Attention-Mamba global modeling.
Authors
Affiliations (3)
Affiliations (3)
- Department of Information Convergence Engineering, Pusan National University, Busan, 46241, Republic of Korea. Electronic address: [email protected].
- Department of Information Convergence Engineering, Pusan National University, Busan, 46241, Republic of Korea; School of Computer Science and Engineering, Pusan National University, Busan, 46241, Republic of Korea. Electronic address: [email protected].
- School of Computer Science and Engineering, Pusan National University, Busan, 46241, Republic of Korea; AI Collaboration Hub Research Center, Pusan National University, Busan, 46241, Republic of Korea. Electronic address: [email protected].
Abstract
Accurately modeling cortical surface geometry from structural MRI is essential for advancing neuroimaging analysis and biomarker discovery. Recent advances in Transformer and Mamba architectures have shown promise in brain surface modeling by effectively capturing complex feature dependencies. However, they typically rely on fixed or absent neighborhood structures, substantially limiting their ability to represent localized cortical variations. Transformers also incur high computational costs, whereas Mamba lacks explicit mechanisms for encoding local spatial context on non-Euclidean domains. To address these limitations and leverage the complementary strengths of both architectures, we propose ASAM (Adaptive Surface representation with Attention and Mamba), a hybrid framework integrating Transformer and Mamba mechanisms within a parallel architecture. ASAM introduces two novel modules designed specifically for cortical surfaces: Dynamic Multi-Hop Selection (DMHS), which adaptively determines optimal receptive fields for each cortical patch, and Dynamic Feature Merging (DFM), which constructs context-sensitive adjacency matrices from geodesic distances and learned affinities. Extensive evaluations on public neuroimaging datasets demonstrate that ASAM achieves superior or highly competitive performance compared to state-of-the-art Transformer- and Mamba-based baselines in CN vs. MCI classification on ADNI and cross-dataset brain age prediction. By dynamically adjusting neighborhood scales, ASAM yields enriched, geometry-aware cortical surface representations, supporting neuroimaging research on risk stratification and developmental brain assessments while motivating further multi-center validation toward clinical translation. Implementation code is publicly available at: https://github.com/sydat2701/ASAM_surface.git.