DyMix: Dynamic Frequency Mixup Scheduler based Unsupervised Domain Adaptation for Enhancing Alzheimer's Disease Identification

TL;DR

This work proposes a novel approach called the dynamic frequency mixup scheduler (DyMix) for unsupervised domain adaptation that dynamically adjusts the magnitude of the frequency regions being mixed from the source and target domains to enhance its generalizability across the target domain.

Abstract

Advances in deep learning (DL)-based models for brain image analysis have significantly enhanced the accuracy of Alzheimer's disease (AD) diagnosis, allowing for more timely interventions. Despite these advancements, most current DL models suffer from performance degradation when inferring on unseen domain data owing to the variations in data distributions, a phenomenon known as domain shift. To address this challenge, we propose a novel approach called the dynamic frequency mixup scheduler (DyMix) for unsupervised domain adaptation. Contrary to the conventional mixup technique, which involves simple linear interpolations between predefined data points from the frequency space, our proposed DyMix dynamically adjusts the magnitude of the frequency regions being mixed from the source and target domains. Such an adaptive strategy optimizes the model's capacity to deal with domain variability, thereby enhancing its generalizability across the target domain. In addition, we incorporate additional strategies to further enforce the model's robustness against domain shifts, including leveraging amplitude-phase recombination to ensure resilience to intensity variations and applying self-adversarial learning to derive domain-invariant feature representations. Experimental results on two benchmark datasets quantitatively and qualitatively validated the effectiveness of our DyMix in that we demonstrated its outstanding performance in AD diagnosis compared to state-of-the-art methods.

Figures (5)

• Figure 1: The primary difference between conventional amplitude mixup techniques and our proposed DyMix. Here, the posterior probabilities in each manipulated image denote the classification accuracy derived from the trained model using their respective augmentation strategies.
• Figure 2: The overall framework of our proposed method consists of two main steps: (i) the pretraining stage for invariant feature representation and (ii) the adaptation stage by dynamic frequency manipulation. This framework ensures a robust approach to learning and adaptation across different domains.
• Figure 3: A t-SNE visualization of (a) the original distribution and (b) the distribution after domain adaptation using our proposed DyMix technique. These visualizations compared the source and target domains across different UDA scenarios, specifically ADNI-1 $\rightarrow$ ADNI-2 (first column) and a ADNI-1 $\rightarrow$ AIBL (second column).
• Figure 4: Illustration of model interpretability using Grad-CAM in the unseen target domain. Here, a red-colored region and a purple-colored region indicate a greater and lesser impact on the model decision, respectively.
• Figure 5: Visual examples of various spatial-based (top right) and frequency-based (bottom right) augmentation methods applied to both source and target brain MRI samples. Each augmentation highlights its unique effect on the image features, illustrating differences in how spatial and frequency components are manipulated to facilitate domain adaptation.