Answers are grounded in this paper and cite their sources.
OBJECTIVE: Clustering is essential for identifying patterns in data. In neuroscience, it is widely used on functional Magnetic Resonance Imaging (fMRI) data to uncover brain activity patterns during cognitive tasks and rest. However, traditional methods like K-means struggle with the high dimensionality and temporal complexity of fMRI data, and often fail to capture spatial or contextual structures in image data, limiting their effectiveness. METHODS: To address these challenges, we propose a deep clustering model that leverages the complementary strengths of Sparse Autoencoder (SAE) and Masked Autoencoder (MAE) for comprehensive feature extraction. The SAE captures global features and reduces data dimensionality while preserving overall structure. In parallel, the MAE reconstructs masked input patches to learn local contextual features-temporal for sequential data like fMRI, and spatial for imagedata. The combination of global and local features yields a more informative representation of the data. These fused features are then fed into a Deep Embedded Clustering (DEC) module, which iteratively refines cluster centers and feature representations using Kullback-Leibler (KL) divergence loss, improving clustering accuracy and robust ness. RESULTS: Experimental results show that our model effectively identifies functional brain activity patterns in task based fMRI and differentiates hippocampal subregions in resting-state fMRI. CONCLUSION: Overall, our approach offers a unified and versatile deep clustering framework capable of handling various data modalities, enabling more accurate and interpretable analysis in both neuroimaging and general high-dimensional data applications. SIGNIFICANCE: The proposed framework explicitly decomposes global and local feature learning, overcoming limitations of single path deep clustering methods and providing physiologically meaningful clustering patterns.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: