Abstract
Deep learning models for Alzheimer’s Disease (AD) identification often suffer from performance degradation when applied to data from different clinical sites, a phenomenon known as domain shift. We present DyMix, a framework that utilizes a dynamic frequency mixup scheduler to bridge the gap between source and target domains. By manipulating the Fourier amplitude spectrum while maintaining phase consistency, DyMix achieves state-of-the-art performance in cross-site AD classification tasks.
Introduction
The identification of structural atrophies in brain MRI scans is critical for diagnosing Alzheimer’s Disease. While deep learning models excel at this task within a single dataset, variations in scanner hardware and imaging protocols across different hospitals lead to "domain shift." This statistical mismatch hinders the deployment of AI tools in diverse clinical settings. Existing Unsupervised Domain Adaptation (UDA) techniques often fail to account for the specific nuances of medical imaging, relying instead on static augmentation strategies.
Methodology
DyMix addresses domain shift by operating in the frequency domain. The framework is built upon two primary innovations:
Fourier-Based Frequency Mixing
Images are decomposed into Amplitude and Phase components using the Fourier Transform. The Amplitude component captures domain-specific textures and intensities, while the Phase component preserves the invariant anatomical structures. DyMix blends the amplitudes of source and target images to create intermediate "bridge" data that helps the model generalize.
Dynamic Scheduler (β)
Unlike standard mixup methods that use fixed ratios, DyMix introduces a tunable β scheduler. This scheduler monitors validation performance and dynamically adjusts the frequency range and magnitude of the mixing process. This adaptivity allows the model to find the optimal alignment for specific dataset pairs, such as transitioning from ADNI-1 to the AIBL study.
Results and Discussion
The DyMix framework was evaluated on benchmark datasets including ADNI and AIBL. In the ADNI-1 to AIBL transfer task, DyMix achieved an accuracy of 91.38%, significantly outperforming traditional UDA methods like DANN and Deep-CORAL.
Visualization using Grad-CAM reveals that the model focuses on key anatomical regions, such as the hippocampus and ventricles, which are primary indicators of AD. This alignment with clinical practice suggests that DyMix not only improves statistical performance but also enhances the biological relevance of the model's decision-making process.
Conclusion
DyMix provides a robust solution for unsupervised domain adaptation in neuroimaging. By leveraging the unique properties of the frequency domain and introducing dynamic scheduling, the framework effectively mitigates the impact of scanner-induced domain shift. Future work will explore the application of DyMix to multi-modal data and other neurodegenerative disorders.
Summary updated