Research map / Detection and segmentation / Weakly supervised detection and segmentation
SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image 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 SGTC, a semi-supervised 3D medical segmentation framework that uses three orthogonal annotated slices per few volumes, CLIP-based semantic-guided auxiliary learning, and triple-view co-training of three V-Net sub-networks.
Problem
Semi-supervised medical image segmentation requires costly slice-by-slice annotation; existing sparse-annotation methods use one or two slices and lack semantic understanding, leading to poor weak-boundary segmentation.
Approach
Combines CLIP text embeddings (via a trainable adapter) that modulate segmentation features with a co-training scheme where three sub-networks each receive supervision from a different pair of orthogonal sparse slices and pseudo-label each other using uncertainty-filtered predictions.
Key results
- On LA2018 with 10% labeled cases, SGTC reaches Dice 0.847 vs 0.784 for BCP (best baseline) using three orthogonal slices.
- On KiTS19 (10% labeled), SGTC achieves Dice 0.933, versus 0.921 for BCP and 0.918 for DTC.
- Ablation on KiTS19 shows SGAL and TVDT each improve Dice over the baseline (0.780 to 0.892 and 0.896), and combined reach 0.915.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SGTC with three orthogonal annotated slices instead of the strongest baseline BCP | Semi-supervised segmentation · LA2018, 10% labeled cases | Dice: 0.784 $\pm$ 0.093 -> 0.847 $\pm$ 0.044 | Helped |
| Use SGTC with three orthogonal annotated slices instead of the strongest baseline BCP | Semi-supervised segmentation · KiTS19, 10% labeled cases | Dice: 0.921 $\pm$ 0.046 -> 0.933 $\pm$ 0.041 | Helped |
| Use SGTC with three orthogonal annotated slices instead of the strongest baseline BCP | Semi-supervised segmentation · LiTS, 10% labeled cases | Dice: 0.922 $\pm$ 0.060 -> 0.927 $\pm$ 0.044 | Helped |
| Add SGAL to a model without SGAL or TVDT (ablation) | SGTC · KiTS19, 9 labeled volumes | Dice: 0.780 -> 0.892 | Helped |
| Add TVDT to a model without SGAL or TVDT (ablation) | SGTC · KiTS19, 9 labeled volumes | Dice: 0.780 -> 0.896 | Helped |
| Add TVDT to a model with SGAL (ablation) | SGTC · KiTS19, 9 labeled volumes | Dice: 0.892 -> 0.915 | Helped |
| Add SGAL to a model with TVDT (ablation) | SGTC · KiTS19, 9 labeled volumes | Dice: 0.896 -> 0.915 | Helped |
| Use the tailored “An image containing the [cls], with the rest being background.” prompt instead of no text prompt (ablation) | SGTC · KiTS19, 6 labeled scans | Dice: 0.873 -> 0.888 | Helped |
| Replace the tailored prompt with “A photo of a [cls].” (ablation) | SGTC · KiTS19, 6 labeled scans | Dice: 0.888 -> 0.882 | Hurt |
| Replace the tailored prompt with the CT/MR prompt (ablation) | SGTC · KiTS19, 6 labeled scans | Dice: 0.888 -> 0.883 | Hurt |
| Fix α at 0.1 instead of using the proposed dynamic coefficient (ablation) | SGTC · KiTS19, 38 labeled scans | Dice: 0.947 -> 0.942 | Hurt |
| Fix α at 0.15 instead of using the proposed dynamic coefficient (ablation) | SGTC · KiTS19, 38 labeled scans | Dice: 0.947 -> 0.944 | Hurt |
Surprising findings
- On KiTS19 with the ablation setting, the TVDT-only variant and the SGAL-only variant perform similarly, and the CAC vs SCA comparison shows only modest gains for most baselines.
Limitations
- Performance degrades when the selected orthogonal slices contain limited foreground information.
- Relies on slices near the volume center containing the target, requiring careful slice selection.
- Results reported with a fixed split on three datasets; comparison of 'SGTC (Dual)' uses different unlabeled counts in some tables.
Experimental setup
Datasets: LA2018 KiTS19 LiTS
Benchmarks: LA2018 KiTS19 LiTS
Metrics: Dice Jaccard 95% Hausdorff distance average surface distance
Baselines: MT UA-MT SASSNet CPS DTC BCP Desco
Builds on: CLIP V-Net Desco BCP DTC
Closest papers by meaning
- Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.88)
- Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.88)
- Integrating Deep Metric Learning with Coreset for Active Learning in 3D Segmentation (NeurIPS 2024 · similarity 0.88)
- Boost the Inference with Co-training: A Depth-guided Mutual Learning Framework for Semi-supervised Medical Polyp Segmentation (CVPR 2025 · similarity 0.87)
- Semi-supervised TEE Segmentation via Interacting with SAM Equipped with Noise-Resilient Prompting (AAAI 2024 · similarity 0.87)
- A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image Segmentation (CVPR 2025 · similarity 0.87)
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.