Functional magnetic resonance imaging (fMRI) denoising is a crucial preprocessing step in neuroimaging studies, as noise degrades the reliability of downstream analyses. Previous approaches for fMRI denoising either rely on predefined noise patterns or train dataset-specific models, restricting their reliability across various datasets due to inter-dataset variations in scanner types, scanning protocols, and preprocessing pipelines. Additionally, applying previous approaches to new datasets requires extensive expert signal/noise annotations. To mitigate this reliance, leveraging existing datasets to train sparsely labeled datasets is a practical solution, but inconsistencies in labeling criteria hinder effective adaptation. To address these challenges, we propose a meta-learning-based semi-supervised domain adaptation framework, enabling the learning of dataset-irrelevant features from sparsely labeled datasets by leveraging existing labeled datasets with two key components: (1) a dataset-irrelevant feature extractor trained by meta-learning to capture noise patterns across multiple datasets, and (2) dataset-specific classifiers optimized by decoupled training to handle inconsistencies in labeling criteria. Our proposed approach shows outstanding performance on four fMRI datasets in both fully labeled and sparsely labeled conditions.