Research map / 3D vision / 3D generation
Learning Dynamic Collaborative Network for Semi-supervised 3D Vessel 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 DiCo, a semi-supervised 3D vessel segmentation framework where a CNN (VNet) and a transformer (UNETR) sub-network dynamically swap teacher-student roles each iteration, plus a multi-view input module and MIP-projected adversarial shape supervision.
Problem
Semi-supervised 3D vessel segmentation with scarce labels; static mean-teacher roles can propagate cognitive biases because the teacher is not always better than the student on complex vessel data.
Approach
Replace the fixed teacher-student roles of mean teacher with dynamic switching based on which sub-network has lower supervised loss on labeled data each iteration, combined with a multi-view (global + local crops) input for the transformer and a discriminator on 2D maximum-intensity projections of predicted masks.
Key results
- On CAS2023 with 5 labeled volumes, DiCo reaches 86.05% DSC, 2.28 points above the second-best MagicNet (83.77 implied by the text's 2.28 gap; table lists MagicNet at 84.13).
- On ImageCAS with 45 labeled volumes, DiCo achieves 73.79% DSC, 74.35% NSD on CAS2023 and 70.93% DSC on Parse2022, the best semi-supervised results on all three datasets for DSC.
- Ablation on CAS2023 shows the base dynamic network improves DSC from 73.94 (MT) to 83.59, with MIP adversarial supervision and multi-view module adding further gains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DiCo instead of MagicNet | semi-supervised vessel segmentation · CAS2023 | DSC (%): 84.13 -> 86.05 | Helped |
| use DiCo instead of MagicNet | semi-supervised vessel segmentation · Parse2022 | DSC (%): 69.19 -> 70.93 | Helped |
| use DiCo instead of SLCNet | semi-supervised vessel segmentation · ImageCAS | DSC (%): 72.78 -> 73.79 | Helped |
| use CNN + ViT rather than CNN + CNN sub-networks (ablation) | DiCo · ImageCAS | DSC (%): 69.45 -> 70.37 | Helped |
| use CNN + ViT rather than ViT + ViT sub-networks (ablation) | DiCo · CAS2023 | DSC (%): 78.56 -> 83.59 | Helped |
| add MIP adversarial supervision without multi-view integration (ablation) | DiCo Base · ImageCAS | DSC (%): 70.37 -> 71.15 | Helped |
| add MIP adversarial supervision without multi-view integration (ablation) | DiCo Base · CAS2023 | DSC (%): 83.59 -> 84.42 | Helped |
| add MIP adversarial supervision without multi-view integration (ablation) | DiCo Base · Parse2022 | DSC (%): 63.85 -> 66.29 | Helped |
| add multi-view integration without MIP adversarial supervision (ablation) | DiCo Base · ImageCAS | DSC (%): 70.37 -> 72.37 | Helped |
| add multi-view integration without MIP adversarial supervision (ablation) | DiCo Base · CAS2023 | DSC (%): 83.59 -> 85.63 | Helped |
| add multi-view integration without MIP adversarial supervision (ablation) | DiCo Base · Parse2022 | DSC (%): 63.85 -> 68.46 | Helped |
| project 3D data to 2D for adversarial supervision (ablation) | DiCo · ImageCAS | DSC (%): 60.79 -> 71.15 | Helped |
Surprising findings
- Projecting 3D masks to 2D via maximum intensity projection for adversarial supervision improved DSC by 17.04, 1.41 and 11.73 points over 3D adversarial supervision across the three datasets.
Limitations
- Relies on 5% labeled data splits; generalization beyond three vessel datasets not shown.
- Multi-view module and dynamic design rely on hand-chosen hyperparameters (n1=2, n2=2, n3=1) and specific CNN-transformer pairing.
- Inference uses only the VNet output, and the transformer's contribution at test time is not used.
Experimental setup
Datasets: ImageCAS CAS2023 Parse2022
Benchmarks: ImageCAS CAS2023 Parse2022
Metrics: DSC NSD ASD
Baselines: MT UA-MT SASSNet SLCNet MagicNet CAML CauSSL GuidedNet VNet CTNet ERNet DSCNet
Builds on: mean teacher VNet UNETR UA-MT
Closest papers by meaning
- vesselFM: A Foundation Model for Universal 3D Blood Vessel Segmentation (CVPR 2025 · similarity 0.87)
- DeNVeR: Deformable Neural Vessel Representations for Unsupervised Video Vessel Segmentation (CVPR 2025 · similarity 0.85)
- Boost the Inference with Co-training: A Depth-guided Mutual Learning Framework for Semi-supervised Medical Polyp Segmentation (CVPR 2025 · similarity 0.83)
- Combinatorial CNN-Transformer Learning with Manifold Constraints for Semi-supervised Medical Image Segmentation (AAAI 2024 · similarity 0.83)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.82)
- DeformCL: Learning Deformable Centerline Representation for Vessel Extraction in 3D Medical Image (CVPR 2025 · similarity 0.82)
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.