Research map / Graph learning / Graph algorithms and clustering
SCCS: Deep Neural Spectral Clustering for Self-Supervised Subcellular Structure Segmentation
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 SCCS, a lightweight GNN-based deep neural spectral clustering framework over MAE-based superpixel graphs that yields consistent subcellular structure segmentation in VEM images without test-time pixel-wise feature extraction or Hungarian synchronization.
Problem
Self-supervised subcellular structure segmentation in volumetric electron microscopy is slow due to test-time pixel-wise feature extraction and requires post-hoc clustering synchronization to keep labels consistent across images.
Approach
Replace per-graph numerical spectral clustering with a learned linear-GCN spectral embedding trained with a Laplacian-based eigendecomposition loss, plus an MLP clustering head supervised by incremental k-means anchored centroids and a smoothness regularizer, operating on superpixel graphs built from MAE features.
Key results
- SCCS reaches average DSC 0.816 on BetaSeg test volume, versus 0.786 for MAESTER and 0.728 for k-means.
- Ablations: removing neural spectral clustering drops average DSC to 0.721; using per-graph k-means with Hungarian matching drops it to 0.673.
- Test-time online inference on a 560x560 image takes about 1.997 s feature extraction, 1.129 s spG construction, and 0.003 s neural spectral clustering; reported as 137.61x faster than MAESTER.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SCCS instead of MAESTER | subcellular segmentation · BetaSeg test | DSC: 0.786 -> 0.816 | Helped |
| use SCCS instead of k-means | subcellular segmentation · BetaSeg test | DSC: 0.728 -> 0.816 | Helped |
| use SCCS instead of DSM | subcellular segmentation · BetaSeg test | DSC: +0.128 | Helped |
| use SCCS instead of DeepCut | subcellular segmentation · BetaSeg test | DSC: +0.150 | Helped |
| use SCCS instead of Han et al. | subcellular segmentation · BetaSeg test | DSC: 0.659 -> 0.816 | Helped |
| add neural spectral clustering instead of numerical spectral clustering (ablation) | SCCS · BetaSeg test | DSC: 0.721 -> 0.816 | Helped |
| add neural clustering with Hungarian matching instead of per-graph k-means (ablation) | SCCS · BetaSeg test | DSC: 0.673 -> 0.816 | Helped |
| add neural clustering instead of incremental k-means (ablation) | SCCS · BetaSeg test | DSC: 0.772 -> 0.816 | Helped |
| set cluster number k to 6 instead of 8 (ablation) | SCCS · BetaSeg test, average | DSC: 0.816 -> 0.559 | Hurt |
| set cluster number k to 7 instead of 8 (ablation) | SCCS · BetaSeg test, average | DSC: 0.816 -> 0.798 | Hurt |
| set cluster number k to 9 instead of 8 (ablation) | SCCS · BetaSeg test, average | DSC: 0.816 -> 0.761 | Hurt |
Surprising findings
- Per-graph k-means with Hungarian matching performs worse than incremental k-means supervision, and the authors report a 0.143 DSC gap relative to the Hungarian variant.
Limitations
- Evaluated only on a single VEM dataset (BetaSeg) with four categories.
- Granule segmentation accuracy remains modest (DSC 0.573), and the number of clusters k must be set manually (k=8).
Experimental setup
Datasets: BetaSeg
Benchmarks: BetaSeg
Metrics: Dice similarity coefficient IoU Sensitivity Accuracy
Baselines: k-means clustering DSM DeepCut Han et al. MAESTER spectral clustering on MAE features and spG
Builds on: masked autoencoder spectral clustering MAESTER DeepCut linear graph convolution
Closest papers by meaning
- MAESTER: Masked Autoencoder Guided Segmentation at Pixel Resolution for Accurate, Self-Supervised Subcellular Structure Recognition (CVPR 2023 · similarity 0.87)
- SpectralNet: Spectral Clustering using Deep Neural Networks (ICLR 2018 · similarity 0.85)
- Automated scalable segmentation of neurons from multispectral images (NeurIPS 2016 · similarity 0.85)
- Efficient Connectivity-Preserving Instance Segmentation with Supervoxel-Based Loss Function (AAAI 2025 · similarity 0.84)
- GraphCL: Graph-based Clustering for Semi-Supervised Medical Image Segmentation (ICML 2025 · similarity 0.84)
- $\epsilon$-Seg: Sparsely Supervised Semantic Segmentation of Microscopy Data (NeurIPS 2025 · 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.