Research map / Label-efficient and robust learning / Semi-supervised learning
Boost the Inference with Co-training: A Depth-guided Mutual Learning Framework for Semi-supervised Medical Polyp 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 RD-Net, a mean-teacher semi-supervised polyp segmentation framework that trains an auxiliary depth-input student for cross-modal mutual learning and depth-guided patch augmentation, while requiring only RGB at inference.
Problem
Semi-supervised polyp segmentation where RGB-D methods need depth input at inference, limiting clinical use; existing semi-supervised methods ignore depth information.
Approach
Train an auxiliary depth-image student alongside the RGB student under mean teacher, use depth-student high-confidence pseudo-labels to guide the RGB student, apply output-level Dice-based contrastive mutual learning across modalities, and select patches for CutMix-style augmentation by depth variance rather than model entropy.
Key results
- On Kvasir-SEG with 2.5% labels (22 images), Dice 86.35% and IoU 80.10%, +3.14% Dice over CML.
- On Kvasir-SEG with 10% labels, Dice 89.07% and IoU 84.21%, versus SupOnly 82.76% Dice.
- On ETIS-Larib with 1/10 labels in the combined setting, Dice improves 28.17% over SupOnly (71.47% vs 43.30% SupOnly Dice).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add depth-guided cross-modal mutual learning (Lud, Lcon and Lcont) to supervised training (ablation) | RD-Net · Kvasir-SEG, 10% labeled | Dice (%): 82.76 -> 88.14 | Helped |
| Add depth-guided patch augmentation to supervised training (ablation) | RD-Net · Kvasir-SEG, 10% labeled | Dice (%): 82.76 -> 88.03 | Helped |
| Add the full set of components, including depth-guided mutual learning and DPA (ablation) | RD-Net · Kvasir-SEG, 10% labeled | Dice (%): 82.76 -> 89.07 | Helped |
| Replace APA with depth-guided patch augmentation (ablation) | RD-Net with AugS1 · Kvasir-SEG, 2.5% labeled | Dice (%): 84.42 -> 86.35 | Helped |
| Replace CutMix with depth-guided patch augmentation (ablation) | RD-Net with AugS1 · Kvasir-SEG, 2.5% labeled | Dice (%): 84.53 -> 86.35 | Helped |
| Replace APA with depth-guided patch augmentation (ablation) | RD-Net with AugS1 · Kvasir-SEG, 10% labeled | Dice (%): 88.50 -> 89.07 | Helped |
| Replace CutMix with depth-guided patch augmentation (ablation) | RD-Net with AugS1 · Kvasir-SEG, 10% labeled | Dice (%): 88.09 -> 89.07 | Helped |
| Vary the pseudo-label confidence threshold γ from 0.7 to 0.95 (ablation) | RD-Net · Kvasir-SEG, 10% labeled | Dice (%): 88.66; 88.74; 89.06; 89.07; 89.03; 88.72 | Mixed |
| Add RD-Net instead of CML | UNet with ResNet-34 · Kvasir-SEG, 2.5% labeled | Dice(%): 83.21 -> 86.35 | Helped |
| Add RD-Net instead of CML | UNet with ResNet-34 · CVC-ClinicDB, 10% labeled | Dice(%): 85.49 -> 87.84 | Helped |
| Add RD-Net instead of AD-MT | UNet with ResNet-34 · ETIS-Larib, 10% labeled, combined training set | Dice(%): 66.29 -> 71.47 | Helped |
Surprising findings
- Patch selection by depth variance outperforms entropy-based APA by 1.93% Dice at 2.5% labels, suggesting model-prediction-based difficulty estimates are unreliable with scarce labels.
Limitations
- Depth maps are generated by a monocular depth estimator (Depth Anything) rather than real depth sensors, so depth quality depends on the estimator.
- Confidence threshold and patch-size choices are hyperparameters tuned on Kvasir-SEG.
Experimental setup
Datasets: Kvasir-SEG CVC-ClinicDB CVC-300 CVC-ColonDB ETIS-Larib
Benchmarks: Kvasir-SEG CVC-ClinicDB CVC-300 CVC-ColonDB ETIS-Larib
Metrics: Dice IoU Acc
Baselines: MT CCT DTC SCP-Net BCP PH-Net CML AD-MT
Builds on: mean teacher CutMix PH-Net adaptive patch augmentation Depth Anything
Closest papers by meaning
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.87)
- Polyper: Boundary Sensitive Polyp Segmentation (AAAI 2024 · similarity 0.87)
- ACL-Net: Semi-supervised Polyp Segmentation via Affinity Contrastive Learning (AAAI 2023 · similarity 0.87)
- Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.85)
- A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image Segmentation (CVPR 2025 · similarity 0.84)
- Combinatorial CNN-Transformer Learning with Manifold Constraints for Semi-supervised Medical Image Segmentation (AAAI 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.