DyMix

This repository provides the official implementation of DyMix, a dynamic frequency Mixup scheduling strategy for unsupervised domain adaptation (UDA) in medical image classification, as described in the following paper:

DyMix: Dynamic Frequency Mixup Scheduler based Unsupervised Domain Adaptation for Enhancing Alzheimer's Disease Prediction
Kwanseok Oh¹, Yooseung Shin^{1, 2}, and Heung-Il Suk¹
(¹Department of Artificial Intelligence, Korea University)
(²Heuron Company Ltd.)

Abstract: Recent advances in deep learning (DL) have substantially improved the accuracy of Alzheimer’s disease (AD) diagnosis from brain images, enabling earlier and more reliable clinical interventions. Nevertheless, most DL-based models often suffer from significant performance degradation when applied to unseen domains owing to variations in data distributions, a challenge commonly referred to as domain shift. To address this issue, we propose DyMix, a dynamic frequency Mixup scheduler for unsupervised domain adaptation (UDA). Built upon a Fourier transformation, DyMix dynamically adjusts the frequency components between source and target domains within selected regions, allowing the model to efficiently capture domain-relevant information. To further enhance robustness, DyMix incorporates intensity-invariant learning and self-adversarial regularization, encouraging the extraction of stable and domain-invariant feature representations. Such an adaptive framework enables robust cross-domain generalization by dynamically aligning domain-specific frequency characteristics while maintaining informative disease-relevant representations. Extensive experiments on two benchmark datasets (i.e., ADNI and AIBL) demonstrate that DyMix consistently outperforms state-of-the-art UDA methods for AD diagnosis. As a result, our method has achieved average performance gains of +6.04% in accuracy and +5.88% in AUC compared to the mean score of all baseline methods across multiple cross-domain scenarios.

Overview

DyMix is designed to address domain shift in multi-site MRI data by dynamically adjusting the frequency mixing region during training. The framework consists of two stages:

Step 1: Pretraining stage

• Coarse intensity augmentation via amplitude–phase recombination (APR)
• Self-adversarial learning for intensity-invariant feature extraction

Step 2: Adaptation stage

• Dynamic frequency Mixup guided by a validation-based scheduler
• Fine-grained frequency alignment between source and target domains

Data Preprocessing Pipeline

This repository provides the complete preprocessing pipeline used in our experiments on the ADNI and AIBL datasets. Due to data usage and licensing restrictions, raw or preprocessed MRI data are not redistributed. Instead, we release all preprocessing scripts and detailed instructions to ensure full reproducibility for researchers with authorized access to the datasets.

Overview

All brain MRI scans from ADNI and AIBL were identically preprocessed using a unified pipeline to minimize domain-specific biases and ensure consistent input representations across datasets.

1. Brain extraction

Non-brain tissues (e.g., skull, neck, and surrounding structures) are removed using HD-BET:

• Tool: HD-BET
• Purpose: Robust skull stripping with minimal manual intervention

hd-bet -i input.nii.gz -o brain.nii.gz -device cpu

2. Linear Registration to MNI152 Space

The skull-stripped images are aligned to the MNI152 standard template using FLIRT from FSL v6.0.1.

• Registration type: Linear (affine)
• Degrees of freedom: 12 (translation, rotation, scale, shear)
• Reference space: MNI152

flirt -in brain.nii.gz \
      -ref MNI152_T1_1mm.nii.gz \
      -out brain_mni.nii.gz \
      -omat affine.mat

This step corrects global linear differences across subjects and datasets.

3. Spatial Normalization and Resampling

All registered volumes are resampled to a uniform spatial resolution of 1 x 1 x 1 mm³

4. Intensity Normalization

To reduce intensity scale variations across scanners and acquisition protocols, Min–Max normalization is applied voxel-wise.

Output Format

Each preprocessed sample is stored as a 3D NIfTI volume with the following properties:

• Shape: 193 x 229 x 193
• Resolution: 1 mm³ isotropic
• Intensity range: [0, 1]
• Coordinate space: MNI152

Data Licensing and Availability

• ADNI (https://adni.loni.usc.edu): data are subject to the ADNI Data Use Agreement and require user registration and approval. Redistribution of raw or derived data is not permitted.
• AIBL (https://aibl.csiro.au): data are distributed under controlled access policies managed by the AIBL study. Users must obtain appropriate authorization prior to data usage.