Research map / Graph learning / Graph self-supervised learning
DCA: Graph-Guided Deep Embedding Clustering for Brain Atlases
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DCA, a graph-guided deep embedding clustering framework that uses a pretrained Swin-UNETR encoder and KNN-graph spectral pseudo-labels with a KL loss to produce individualized and group voxel-wise brain atlases, evaluated on HCP/ABIDE/ADNI.
Problem
Existing brain atlases are predefined group-level templates with fixed granularity and coarse anatomical coverage, and off-the-shelf clustering ignores spatial contiguity and struggles with low-SNR voxel-level fMRI.
Approach
Combine a masked-reconstruction pretrained Swin-UNETR encoder with learnable centroids trained by KL divergence to graph-cut (spectral) pseudo-labels computed on a 26-neighbour voxel graph, yielding spatially contiguous, functionally homogeneous voxel-wise parcellations at arbitrary resolution and ROI.
Key results
- DCA improves homogeneity by 98.8% and silhouette by 29% on average across 41 to 800 parcels versus existing atlases (100 HCP subjects).
- At 200 parcels, DCA improves homogeneity by 77.7% and silhouette by 19.5% over the Schaefer baseline.
- Task-adapted atlas DCA^gender_100 reaches 0.82 accuracy vs 0.73 for Watershed (100) on CNN gender classification.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCA instead of Schaefer at matched resolutions | DCA group atlas · HCP, 41–800 parcels | homogeneity: 98.8% | Helped |
| Use DCA instead of Schaefer at matched resolutions | DCA group atlas · HCP, 41–800 parcels | silhouette: 29% | Helped |
| Fine-tune Swin-UNETR for gender classification to create a task-specific atlas | DCA100 (individual) · CNN gender classification, HCP | accuracy: 0.70 -> 0.82 | Helped |
| Fine-tune Swin-UNETR for gender classification to create a task-specific atlas | DCA100 (individual) · k-GNN gender classification, HCP | accuracy: 0.725 -> 0.825 | Helped |
| Add pretrained Swin-UNETR embeddings instead of clustering raw fMRI (ablation) | DCA · HCP, homogeneity | homogeneity: 0.0740±0.0186 -> 0.1004±0.0216 | Helped |
| Add pretrained Swin-UNETR embeddings instead of clustering raw fMRI (ablation) | DCA · HCP, silhouette | silhouette: 0.0061±0.0045 -> 0.0304±0.0068 | Helped |
| Add graph cut to K-Means on raw fMRI (ablation) | fMRI clustering · HCP, homogeneity | homogeneity: 0.0740±0.0186 -> 0.0860±0.0201 | Helped |
| Add graph cut to K-Means on raw fMRI (ablation) | fMRI clustering · HCP, silhouette | silhouette: 0.0061±0.0045 -> 0.0179±0.0053 | Helped |
| Add graph cut to K-Means on embeddings (ablation) | embedding clustering · HCP, homogeneity | homogeneity: 0.0790±0.0203 -> 0.0897±0.0210 | Helped |
| Add graph cut to K-Means on embeddings (ablation) | embedding clustering · HCP, silhouette | silhouette: 0.0154±0.0073 -> 0.0208±0.0061 | Helped |
| Add KL-guided joint optimization of encoder parameters and centroids (ablation) | DCA with embedding graph cut · HCP, homogeneity | homogeneity: 0.0897±0.0210 -> 0.1004±0.0216 | Helped |
| Add KL-guided joint optimization of encoder parameters and centroids (ablation) | DCA with embedding graph cut · HCP, silhouette | silhouette: 0.0208±0.0061 -> 0.0304±0.0068 | Helped |
Surprising findings
- Without KL-guided joint optimization, the graph-cut and K-means variants yield far lower homogeneity and many more connected components per parcel.
- Reproducibility with random seeds is only moderate (Dice about 0.8 across runs) unless both seeds are fixed.
Limitations
- Voxel-wise representation learning and clustering incur substantial memory and computational cost at whole-brain scale.
- Fixed KNN graphs may attenuate long-range functional relationships.
- Only single-modality fMRI is used.
Experimental setup
Datasets: HCP ABIDE I ADNI CHCP
Benchmarks: AtlaScore HCP gender HCP fluid intelligence HCP task classification (7 and 24-way) ABIDE ASD ADNI AD/MCI
Metrics: homogeneity silhouette coefficient DCBC classification accuracy F1
Baselines: Yeo Brodmann Schaefer AAL MUSE MMP GIANT Watershed Allen Human Reference Atlas K-Means Graph Cut
Builds on: Swin-UNETR deep clustering with KL target distribution (DEC) spectral clustering Hungarian algorithm
Closest papers by meaning
- A Brain Graph Foundation Model: Pre-Training and Prompt-Tuning across Broad Atlases and Disorders (ICLR 2026 · similarity 0.87)
- Geometry-Guided Generative Representation for Functional Brain Graphs (ICML 2026 · similarity 0.86)
- MAGNET: Manifold-Aware Graph Diffusion Network for Connectome Generation (NeurIPS 2026 · similarity 0.85)
- Learning to cluster neuronal function (NeurIPS 2025 · similarity 0.85)
- FlatClip: Reusing Image Foundation Models for fMRI Representation Learning via Cortical Flatmaps (NeurIPS 2026 · similarity 0.84)
- Unsupervised Representation Learning of Brain Activity via Bridging Voxel Activity and Functional Connectivity (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.