Frequency Mixup Manipulation Based Unsupervised Domain Adaptation for Brain Disease Identification

Unsupervised Domain Adaptation (UDA), which transfers the learned knowledge from a labeled source domain to an unlabeled target domain, has been widely utilized in various medical image analysis approaches. Recent advances in UDA have shown that manipulating the frequency domain between source and target distributions can significantly alleviate the domain shift problem. However, a potential drawback of these methods is the loss of semantic information in the low-frequency spectrum, which can make it difficult to consider semantic information across the entire frequency spectrum. To deal with this problem, we propose a frequency mixup manipulation that utilizes the overall semantic information of the frequency spectrum in brain disease identification. In the first step, we perform self-adversarial disentangling based on frequency manipulation to pretrain the model for intensity-invariant feature extraction. Then, we effectively align the distributions of both the source and target domains by using mixed-frequency domains. In the extensive experiments on ADNI and AIBL datasets, our proposed method achieved outstanding performance over other UDA-based approaches in medical image classification. Code is available at: https://github.com/ku-milab/FMM .