Research map / Video / Video generation
Label-Efficient Data Augmentation with Video Diffusion Models for Guidewire Segmentation in Cardiac Fluoroscopy
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 SF-VD, a three-stage pair of 2D diffusion models with frame-consistency sampling and segmentation-guided contrast control that synthesizes labeled cardiac fluoroscopy videos to augment guidewire segmentation training.
Problem
Guidewire segmentation in cardiac fluoroscopy videos needs large annotated datasets, but annotating thin wires is labor-intensive and few labeled videos exist.
Approach
Decomposes video distribution into a scene distribution (a mask-conditioned 2D DM trained on partially annotated images) and a motion distribution (a 2D DM generating frames conditioned on neighboring frames and masks), sampled via frame-consistency ordering, plus a segmentation-network classifier-style guidance to vary wire contrast.
Key results
- SF-VD achieves the lowest FVD (24.38) versus VD (56.49) and LVDM (27.55) on synthesized fluoroscopy video quality.
- With SF-VD augmentation, Dice improves to 0.61 for Dynvnet, SwinUnetr and MLP-Mixer, and to 0.66 for RTGN (from 0.49 without augmentation).
- Ablation shows frame-consistency mainly improves Dice, sensitivity and precision, while segmentation guidance mainly reduces HD and G2RE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SF-VD augmentation instead of the strongest competing augmentation | Dynvnet · Guidewire segmentation test | Dice: 0.58 -> 0.61 | Helped |
| Add SF-VD augmentation instead of the strongest competing augmentation | Dynvnet · Guidewire segmentation test | HD: 4.83 -> 3.93 | Helped |
| Add SF-VD augmentation instead of the strongest competing augmentation | SwinUnetr · Guidewire segmentation test | Dice: 0.58 -> 0.61 | Helped |
| Add SF-VD augmentation instead of the strongest competing augmentation | MLP-Mixer · Guidewire segmentation test | Dice: 0.60 -> 0.61 | Helped |
| Add SF-VD augmentation instead of the strongest competing augmentation | PARN · Guidewire segmentation test | Dice: 0.60 -> 0.62 | Helped |
| Add SF-VD augmentation instead of the strongest competing augmentation | RTGN · Guidewire segmentation test | Dice: 0.62 -> 0.66 | Helped |
| Add SF-VD augmentation instead of the strongest competing augmentation | TTN · Guidewire segmentation test | Dice: 0.52 -> 0.58 | Helped |
| Add frame-consistency sampling (ablation) | SF-VD · Guidewire segmentation | Dice | Helped |
| Add frame-consistency sampling (ablation) | SF-VD · Guidewire segmentation | HD | No clear effect |
| Add segmentation-guided contrast modulation (ablation) | SF-VD · Guidewire segmentation | HD | Helped |
| Add segmentation-guided contrast modulation (ablation) | SF-VD · Guidewire segmentation | Dice | No clear effect |
| Combine frame-consistency sampling and segmentation-guided contrast modulation (ablation) | SF-VD · Guidewire segmentation | Dice | Helped |
Surprising findings
- LVDM attains comparable FVD to SF-VD and even slightly better HD for some networks, so SF-VD's advantage is mainly in diversity and downstream segmentation.
- Generic augmentation methods such as geometric augmentation and pre-training (CL, MVI) give inconsistent or limited gains over no augmentation.
Limitations
- Dual DDPM architecture is inefficient, taking about 40 seconds per 16-frame video on an A100.
- Relies on existing wire mask sequences, limiting variability in wire motion.
- Limited number of training videos may affect quality and variability of generated outputs.
Experimental setup
Datasets: F-video (400 fully annotated fluoroscopy videos) P-image (409 partially annotated fluoroscopy videos; 4000 annotated and 10000 unannotated frames)
Benchmarks: F-video test split (10%)
Metrics: FVD Diversity Score Overfitting Score Dice Hausdorff Distance G2RE R2GE sensitivity precision
Baselines: Geometric augmentation Contrastive Learning Masked Volume Modeling Video Diffusion Model (VD) LVDM V-Net Dynvnet SwinUnetr MLP-Mixer PARN RTGN TTN
Builds on: IDDPM classifier-free diffusion guidance classifier guidance Video Diffusion Model LVDM
Closest papers by meaning
- Noise-Consistent Siamese-Diffusion for Medical Image Synthesis and Segmentation (CVPR 2025 · similarity 0.82)
- DeNVeR: Deformable Neural Vessel Representations for Unsupervised Video Vessel Segmentation (CVPR 2025 · similarity 0.81)
- vesselFM: A Foundation Model for Universal 3D Blood Vessel Segmentation (CVPR 2025 · similarity 0.80)
- VidSeg: Training-free Video Semantic Segmentation based on Diffusion Models (CVPR 2025 · similarity 0.80)
- DiffusionBlend: Learning 3D Image Prior through Position-aware Diffusion Score Blending for 3D Computed Tomography Reconstruction (NeurIPS 2024 · similarity 0.80)
- MedSegDiff-V2: Diffusion-Based Medical Image Segmentation with Transformer (AAAI 2024 · similarity 0.80)
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.