Research map / Label-efficient and robust learning / Domain adaptation and generalization
Alleviate and Mining: Rethinking Unsupervised Domain Adaptation for Mitochondria Segmentation from Pseudo-Label Perspective
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 R4MITO, a UDA framework for mitochondria segmentation in EM images that combines reliable prototype pseudo-labels with correlation-wise and rank-aware agent-level consistency regularization.
Problem
Unsupervised domain adaptation for mitochondria segmentation from EM images, where pseudo-labels are unreliable under domain shift and pixel-wise consistency underuses supervision signals.
Approach
Builds target-domain prototypes from high-confidence features initialized by source prototypes, and enforces consistency on agent-level inter-pixel correlations plus agent-ranking permutation probabilities (top-4 agents) under weak/strong perturbations.
Key results
- On VNC III→Lucchi (Subset1), R4MITO reaches 73.7% IoU, 1.9 points above CAFA (71.8%).
- On VNC III→Lucchi (Subset2), IoU reaches 78.6%, a 3.2-point gain over CAFA.
- On MitoEM-R→MitoEM-H and MitoEM-H→MitoEM-R, IoU reaches 77.3% and 81.8%, improving on CAFA by 1.0 and 1.2 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reliable prototype pseudo-labels (ablation) | CAFA baseline · VNC III → Lucchi (Subset1) | IoU(%): 71.8 -> 72.9 | Helped |
| add correlation-wise consistency regularization (ablation) | CAFA baseline · VNC III → Lucchi (Subset1) | IoU(%): 71.8 -> 72.2 | Helped |
| add reliable prototype pseudo-labels and correlation-wise consistency regularization (ablation) | CAFA baseline · VNC III → Lucchi (Subset1) | IoU(%): 71.8 -> 72.8 | Helped |
| add rank-aware relationship consistency regularization (ablation) | R4MITO without rank-aware regularization · VNC III → Lucchi (Subset1) | IoU(%): +1.0% | Helped |
| use TopK agent selection instead of selecting all agents (ablation) | R4MITO · VNC III → Lucchi (Subset1) | IoU(%): 73.4 -> 73.7 | Helped |
| select agents randomly instead of using TopK selection (ablation) | R4MITO · VNC III → Lucchi (Subset1) | IoU(%): 73.7 -> 73.1 | Hurt |
| set agent number K to 32 rather than 128 (ablation) | R4MITO · VNC III → Lucchi (Subset1) | IoU(%): 73.7 -> 73.3 | Hurt |
| set agent number K to 64 rather than 128 (ablation) | R4MITO · VNC III → Lucchi (Subset1) | IoU(%): 73.7 -> 73.5 | Hurt |
| set agent number K to 256 rather than 128 (ablation) | R4MITO · VNC III → Lucchi (Subset1) | IoU(%): 73.7 -> 73.6 | Hurt |
| set λpro to 0.05 rather than 0.1 (ablation) | R4MITO · VNC III → Lucchi (Subset1) | IoU(%): 73.7 -> 73.3 | Hurt |
| set λrank to 0.05 rather than 0.1 (ablation) | R4MITO · VNC III → Lucchi (Subset1) | IoU(%): 73.7 -> 73.4 | Hurt |
| set reliable target feature threshold τ to 0.5 rather than 0.7 (ablation) | R4MITO · VNC III → Lucchi (Subset1) | IoU(%): 73.7 -> 73.2 | Hurt |
Limitations
- Rank permutations restricted to top-4 agents as a computational shortcut.
- Threshold tau and loss weights require empirical tuning.
Experimental setup
Datasets: VNC III Lucchi MitoEM
Benchmarks: Lucchi MitoEM
Metrics: mAP F1 MCC IoU
Baselines: Advent DAMT-Net UALR DA-VSN DA-ISC SAPAN CAFA DANN AdaptSegNet Y-Net
Builds on: CAFA Advent DA-ISC Mean Teacher-style weak/strong consistency
Closest papers by meaning
- Evidential Uncertainty-Guided Mitochondria Segmentation for 3D EM Images (AAAI 2024 · similarity 0.89)
- Electron Microscopy Images as Set of Fragments for Mitochondrial Segmentation (AAAI 2024 · similarity 0.87)
- RankMatch: Exploring the Better Consistency Regularization for Semi-supervised Semantic Segmentation (CVPR 2024 · similarity 0.86)
- Vox-UDA: Voxel-wise Unsupervised Domain Adaptation for Cryo-Electron Subtomogram Segmentation with Denoised Pseudo-Labeling (AAAI 2025 · similarity 0.85)
- $\epsilon$-Seg: Sparsely Supervised Semantic Segmentation of Microscopy Data (NeurIPS 2025 · similarity 0.84)
- MATCH: Multi-faceted Adaptive Topo-Consistency for Semi-Supervised Histopathology Segmentation (NeurIPS 2025 · similarity 0.83)
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.