Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-supervised TEE Segmentation via Interacting with SAM Equipped with Noise-Resilient Prompting
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 a semi-supervised LAA segmentation method for TEE images that uses SAM, prompted with noise-resilient points from a self-reconstruction teacher-student network, to refine pseudo labels, plus feature-level self-perturbation.
Problem
Semi-supervised left atrial appendage segmentation in transesophageal echocardiography (TEE) suffers from pseudo-label bias due to distribution mismatch between highly imbalanced labeled and unlabeled data, and from scarce annotations and noisy TEE images.
Approach
Incorporates the foundation model SAM into a teacher-student SSL framework: a self-reconstruction (denoising) module produces reconstructed images and candidate masks from which consensus high-confidence points serve as noise-resilient SAM prompts, and SAM outputs are adaptively fused with the teacher's predictions into pseudo labels; a self-perturbation module (SPFA) swaps encoder features betwee
Key results
- With 20% labeled data, the method reaches Dice 0.8604 and ASD 1.65, versus 0.8405 for the best baseline BCP and 0.8290 for fully supervised Unet.
- With 50% labeled data, Dice 0.8830 and ASD 1.28, outperforming UniMatch (0.8764) and ST++ (0.8694).
- Removing SAM-powered pseudo label generation drops 20%-label Dice from 0.8604 to 0.8366.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the SAM-powered pseudo label generator (GPL) (ablation) | semi-supervised Unet · TEE dataset, 20% labeled | Dice: 0.8293 -> 0.8469 | Helped |
| add the SAM-powered pseudo label generator (GPL) (ablation) | semi-supervised Unet · TEE dataset, 20% labeled | ASD: 2.08 -> 1.92 | Helped |
| add SPFA (ablation) | semi-supervised Unet · TEE dataset, 20% labeled | Dice: 0.8293 -> 0.8366 | Helped |
| add SPFA (ablation) | semi-supervised Unet · TEE dataset, 20% labeled | ASD: 2.08 -> 1.94 | Helped |
| add the self-reconstruction mechanism (SRM) for prompts and reconstructed inputs (ablation) | SAM-powered semi-supervised Unet · TEE dataset, 20% labeled | Dice: 0.8249 -> 0.8604 | Helped |
| add the self-reconstruction mechanism (SRM) for prompts and reconstructed inputs (ablation) | SAM-powered semi-supervised Unet · TEE dataset, 20% labeled | ASD: 2.16 -> 1.65 | Helped |
| use SPFA instead of DA (ablation) | GPL-based semi-supervised Unet · TEE dataset, 20% labeled | Dice: 0.8385 -> 0.8604 | Helped |
| use SPFA instead of DA (ablation) | GPL-based semi-supervised Unet · TEE dataset, 20% labeled | ASD: 1.97 -> 1.65 | Helped |
| use SPFA instead of FA dropout (ablation) | GPL-based semi-supervised Unet · TEE dataset, 20% labeled | Dice: 0.8404 -> 0.8604 | Helped |
| use SPFA instead of FA dropout (ablation) | GPL-based semi-supervised Unet · TEE dataset, 20% labeled | ASD: 1.84 -> 1.65 | Helped |
Surprising findings
- Directly using SAM to generate pseudo labels without the self-reconstruction prompts yields Dice 0.8249, worse than the fully supervised Unet, indicating large domain gap for SAM on TEE.
Limitations
- Evaluated only on a single in-house TEE dataset of 100 patients.
- Relies on SAM's zero-shot quality on TEE, which the paper shows is weak without the proposed prompting.
Experimental setup
Datasets: in-house TEE dataset (100 patients)
Benchmarks: in-house TEE LAA dataset
Metrics: Dice ASD
Baselines: Unet ST++ UniMatch MCF BCP UCMT
Builds on: Mean Teacher Neighbor2Neighbor Segment Anything UniMatch ST++
Closest papers by meaning
- Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.88)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.87)
- Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation (CVPR 2024 · similarity 0.87)
- Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- beta-FFT: Nonlinear Interpolation and Differentiated Training Strategies for Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- Combinatorial CNN-Transformer Learning with Manifold Constraints for Semi-supervised Medical Image Segmentation (AAAI 2024 · 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.