Research map / Label-efficient and robust learning / Domain adaptation and generalization
Vox-UDA: Voxel-wise Unsupervised Domain Adaptation for Cryo-Electron Subtomogram Segmentation with Denoised Pseudo-Labeling
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 Vox-UDA, a voxel-wise unsupervised domain adaptation framework for cryo-ET subtomogram segmentation that combines a frequency-based noise generation module and a gradient-guided improved bilateral filter for denoised pseudo-labeling.
Problem
Cryo-ET subtomogram segmentation needs manual labels; simulated source data have fixed noise levels and different macromolecules than real target data, causing noise mismatch and domain shift when adapting to real cryo-ET data.
Approach
Generates target-like Gaussian noise (variance estimated from a sampled subset of target data after high-pass Fourier filtering) and adds it to source inputs, and denoises target inputs with an improved bilateral filter whose range kernel uses Laplacian-based gradient differences rather than intensity differences, to produce cleaner teacher pseudo-labels.
Key results
- On Poly-GA with source 1bxn/1f1b/1yg6, Vox-UDA (w IBF) reaches 50.3 mIoU and 65.9 Dice, versus 40.4 mIoU and 56.8 Dice for the best prior UDA baseline LE-UDA-level comparison (ASC/LE-UDA range).
- With source 2byu/2h12/21db, Vox-UDA (w IBF) achieves 49.5 mIoU and 65.2 Dice, outperforming LE-UDA (43.1 mIoU, 58.6 Dice).
- Ablations show removing NGM drops mIoU to 41.9 and removing pseudo-labels drops it to 46.2 versus 50.3 for the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Vox-UDA with IBF instead of LE-UDA | VoxResNet UDA · Poly-GA (source: 1bxn, 1f1b, 1yg6) | mIoU: 40.4 → 50.3 | Helped |
| Use Vox-UDA with IBF instead of LE-UDA | VoxResNet UDA · Poly-GA (source: 1bxn, 1f1b, 1yg6) | Dice: 56.8 → 65.9 | Helped |
| Use Vox-UDA with IBF instead of LE-UDA | VoxResNet UDA · Poly-GA (source: 2byu, 2h12, 21db) | mIoU: 43.1 → 49.5 | Helped |
| Add NGM (ablation) | Vox-UDA · Poly-GA (source: 1bxn, 1f1b, 1yg6) | mIoU: 41.9 → 50.3 | Helped |
| Add pseudo-labeling (ablation) | Vox-UDA · Poly-GA (source: 1bxn, 1f1b, 1yg6) | mIoU: 46.2 → 50.3 | Helped |
| Replace NGM denoising with Bilateral Filter denoising (ablation) | Vox-UDA pseudo-labeling · Poly-GA (source: 1bxn, 1f1b, 1yg6) | mIoU: 48.5 → 49.1 | Helped |
| Replace Bilateral Filter with improved Bilateral Filter (ablation) | Vox-UDA pseudo-labeling · Poly-GA (source: 1bxn, 1f1b, 1yg6) | mIoU: 49.1 → 50.3 | Helped |
| Set N_sample to 10 rather than 5, 15, or 20 (ablation) | Vox-UDA · Poly-GA (source: 1bxn, 1f1b, 1yg6) | mIoU: 41.2; 50.3; 43.6; 42.6 | Mixed |
| Set the high-pass filter rate rho to 24.4% rather than other tested rates (ablation) | Vox-UDA · Poly-GA (source: 1bxn, 1f1b, 1yg6) | mIoU: 41.7; 43.5; 50.3; 41.0 | Mixed |
| Use consistency-loss weights [0.2, 0.2, 0.3, 0.3] (ablation) | Vox-UDA · Poly-GA (source: 1bxn, 1f1b, 1yg6) | mIoU: 44.4; 50.3; 45.3; 42.2 | Mixed |
| Set sigma_d to 120 (ablation) | Vox-UDA IBF · Poly-GA (source: 1bxn, 1f1b, 1yg6) | mIoU: 49.3; 50.3; 49.1; 47.0 | Mixed |
| Set sigma_r to 1.2 (ablation) | Vox-UDA IBF · Poly-GA (source: 1bxn, 1f1b, 1yg6) | mIoU: 47.5; 48.0; 50.3; 49.5 | Mixed |
Surprising findings
- Vox-UDA slightly exceeds the fully supervised VoxResNet upper bound on several metrics (e.g., 50.3 vs 46.0 mIoU).
Limitations
- Only evaluated on one real dataset (Poly-GA) with volumes resized to 32^3, limiting scalability claims.
- Key hyperparameters (sampled count, filter rate, loss weights, sigma values) selected by sweeping on the evaluation data, with sensitivity to them.
- Simulated source data limited to six macromolecules; target and source macromolecule categories differ.
Experimental setup
Datasets: simulated cryo-ET subtomograms (1bxn, 1f1b, 1yg6; 2byu, 2h12, 21db) Poly-GA
Benchmarks: Poly-GA
Metrics: mIoU Dice
Baselines: DANN PDAM ASC LE-UDA Fully Supervised VoxResNet w/o adaptation
Builds on: VoxResNet DANN mean teacher bilateral filter
Closest papers by meaning
- cryoSPHERE: Single-Particle HEterogeneous REconstruction from cryo EM (ICLR 2025 · similarity 0.87)
- BOE-ViT: Boosting Orientation Estimation with Equivariance in Self-Supervised 3D Subtomogram Alignment (CVPR 2025 · similarity 0.87)
- CryoFM: A Flow-based Foundation Model for Cryo-EM Densities (ICLR 2025 · similarity 0.86)
- CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM (NeurIPS 2024 · similarity 0.86)
- CryoSPIN: Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference (NeurIPS 2024 · similarity 0.86)
- Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery (NeurIPS 2025 · similarity 0.86)
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.