Search

For the machine learning-based prediction of the conversion from mild cognitive impairment to Alzheimer's disease, the collection of sufficient data to train a model is required, which involves a lot of time and expense. When data is not enough, combining public and in-house data may be appropriate by applying domain adaptation that alleviates inter-site heterogeneity. Existing methods simultaneously transform in-house and public data to represent them into a common feature space, and then train a classifier using labels in public data. However, this procedure causes the time- and cost-consuming re-training of classifier whenever in-house data changes, and also inheres the risk of information loss in public data. Motivated by this, we propose a method that only transforms in-house data while preserving public data, namely one-way domain adaptation. The proposed method represents in-house data similar with public data by matching the data distribution and the connectivity between brain regions with mean vectors and covariance matrices, respectively. Then, the pre-trained classifier in public data is applied to predict AD conversion for in-house data. The experiments, which use the Australian Imaging Biomarkers and Lifestyle Study of Aging and the Open Access Series of Imaging Studies as the in-house data and the Alzheimer's Disease Neuroimaging Initiative as the public data, show the effectiveness and efficiency of the proposed method, improving prediction performance about 34.8% on average without labels in the in-house datasets. [ABSTRACT FROM AUTHOR]

• We tackle a realistic problem setting of domain adaptation, where most domains are label-deficient and need to be helped and recent data become more sparsely labeled which makes the learning even more difficult. • To tackle this problem, we propose a mutual domain adaptation, which transfer label information both-way, to search a common feature space that matches different data distributions, preserves original manifolds, and maximize consistency between labeled samples with pseudo-labeling via semi-supervised learning. • We validate the proposed method on benchmark datasets for domain adaptation with varying the rate of labels, on which it outperforms relevant baselines and is especially better for the sparsely labeled data so as to be suitable for real-world scenarios. To solve the label sparsity problem, domain adaptation has been well-established, suggesting various methods such as finding a common feature space of different domains using projection matrices or neural networks. Despite recent advances, domain adaptation is still limited and is not yet practical. The most pronouncing problem is that the existing approaches assume source-target relationship between domains, which implies one domain supplies label information to another domain. However, the amount of label is only marginal in real-world domains, so it is unrealistic to find source domains having sufficient labels. Motivated by this, we propose a method that allows domains to mutually share label information. The proposed method finds a projection matrix that matches the respective distributions of different domains, preserves their respective geometries, and aligns their respective class boundaries. The experiments on benchmark datasets show that the proposed method outperforms relevant baselines. In particular, the results on varying proportions of labels present that the fewer labels the better improvement. [ABSTRACT FROM AUTHOR]