DMFLN: A dynamic multi-scale focus learning framework for Alzheimer's disease classification

Magnetic resonance imaging (MRI) of gray matter plays a crucial role in the diagnosis of Alzheimer's disease (AD). Recent advances in multiscale learning techniques have improved AD classification by capturing structural information at multiple scales. However, effectively balancing the contributions of these multiscale features remains a significant challenge.

To address this issue, we propose a Dynamic Multiscale Feature Learning Network (DMFLN) for AD classification. The framework incorporates a pyramid self-attention mechanism to capture high-level global contextual features and model long-range dependencies. Additionally, a residual wavelet transform is utilized to extract fine-grained local structural features. The DMFLN adaptively adjusts the weights of features across different scales, enabling a balanced fusion of global topological representations and local morphological details.

We evaluate our approach on T1-weighted MRI scans from the ADNI dataset. The proposed method achieves classification accuracies of 96.32% ± 0.51%, 94.62% ± 0.39%, and 93.07% ± 0.81% for AD vs. NC, AD vs. MCI, and NC vs. MCI tasks, respectively.

Compared to state-of-the-art approaches, the DMFLN framework offers improved performance by effectively addressing the challenge of multiscale feature weighting, which is often a bottleneck in multiscale fusion-based AD classification.

The DMFLN framework demonstrates significant improvements in AD classification by adaptively integrating global and local structural information from gray matter. These results highlight the potential of dynamic multiscale feature learning in advancing neuroimaging-based AD diagnosis.