Research map / Probabilistic and scientific modeling / Molecules and proteins
Blood Flow Speed Estimation with Optical Coherence Tomography Angiography Images
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 OCTA-Flow, a Swin-based encoder-decoder that regresses pixel-wise blood flow speed directly from OCTA images, using ODT measurements as noisy pseudo labels, with an Adaptive Window Fusion module and a CRF decoder.
Problem
Blood flow speed estimation in the mouse cerebral cortex is hard: ODT is costly, needs complex hardware and signal processing, and suffers from angle and motion artifacts, while no ground-truth flow data exists.
Approach
Learn to predict flow speed directly from routinely acquired OCTA vascular images, supervised by ODT data as pseudo labels, with multiscale window-attention fusion (AWF) conditioned on input and a hierarchical neural CRF decoder that smooths predictions and suppresses ODT artifacts.
Key results
- On the Anesthetized Dataset, OCTA-Flow reaches Abs Rel 0.353 and RMSE 6.278, beating the second-best method by 0.009 Abs Rel and 0.150 RMSE.
- On the Awake Dataset, OCTA-Flow reaches Abs Rel 0.318 and RMSE 6.037, improving over the second-best by 0.041 Abs Rel and 0.278 RMSE.
- Ablation on the Anesthetized Dataset: Base+AWF+CRFD reduces Abs Rel from 0.393 to 0.328 relative to the UNet-like base.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OCTA-Flow instead of the strongest baseline | Blood flow estimation · Anesthetized Dataset | Abs Rel: 0.362 -> 0.353 | Helped |
| Use OCTA-Flow instead of the strongest baseline | Blood flow estimation · Awake Dataset | Abs Rel: 0.359 -> 0.318 | Helped |
| Add AWF to the baseline (ablation) | UNet-like baseline on Anesthetized Dataset · Anesthetized Dataset | Abs Rel: 0.393 -> 0.344 | Helped |
| Add CRFD to the baseline (ablation) | UNet-like baseline on Anesthetized Dataset · Anesthetized Dataset | Abs Rel: 0.393 -> 0.358 | Helped |
| Add both AWF and CRFD to the baseline (ablation) | UNet-like baseline on Anesthetized Dataset · Anesthetized Dataset | Abs Rel: 0.393 -> 0.328 | Helped |
| Replace AWF with pyramid pooling (ablation) | OCTA-Flow on Anesthetized Dataset · Anesthetized Dataset | Abs Rel: 0.328 -> 0.342 | Hurt |
| Use one window attention block (ablation) | AWF on Anesthetized Dataset · Anesthetized Dataset | Abs Rel: 0.358 -> 0.348 | Helped |
| Add a second window attention block with a different window size (ablation) | AWF on Anesthetized Dataset · Anesthetized Dataset | Abs Rel: 0.348 -> 0.343 | Helped |
| Add a third window attention block with a different window size (ablation) | AWF on Anesthetized Dataset · Anesthetized Dataset | Abs Rel: 0.343 -> 0.335 | Helped |
| Add dynamic gate-generated weights to multiscale window fusion (ablation) | AWF on Anesthetized Dataset · Anesthetized Dataset | Abs Rel: 0.335 -> 0.328 | Helped |
| Use one HiCRF block (ablation) | CRFD on Anesthetized Dataset · Anesthetized Dataset | Abs Rel: 0.344 -> 0.339 | Helped |
| Add a second cascaded HiCRF block (ablation) | CRFD on Anesthetized Dataset · Anesthetized Dataset | Abs Rel: 0.339 -> 0.334 | Helped |
Surprising findings
- Diffusion-based depth models (Diff Depth, ECoDepth) perform worse than Swin-based regression models on this flow task, with ECoDepth reaching Abs Rel 0.766 on the Awake Dataset.
Limitations
- Uses ODT measurements as noisy pseudo labels, so the model may inherit ODT artifacts and systematic errors.
- Evaluated only on mouse cerebral cortex data from two conditions (anesthetized and awake).
Experimental setup
Datasets: Anesthetized Dataset (OCTA-ODT pairs, mouse cerebral cortex) Awake Dataset (OCTA-ODT pairs, mouse cerebral cortex)
Benchmarks: Anesthetized Dataset Awake Dataset
Metrics: Abs Rel RMSE
Baselines: BTS IEBins NeuWin Ord Ent Diff Depth ECoDepth
Builds on: Swin Transformer neural CRF scale-invariant logarithmic loss
Closest papers by meaning
- MuTri: Multi-view Tri-alignment for OCT to OCTA 3D Image Translation (CVPR 2025 · similarity 0.86)
- EchoXFlow: A Beamspace Echocardiography Dataset for Cardiac Motion, Flow, and Function (NeurIPS 2026 · similarity 0.82)
- GaraMoSt: Parallel Multi-Granularity Motion and Structural Modeling for Efficient Multi-Frame Interpolation in DSA Images (AAAI 2025 · similarity 0.81)
- VasTSD: Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis (CVPR 2025 · similarity 0.81)
- vesselFM: A Foundation Model for Universal 3D Blood Vessel Segmentation (CVPR 2025 · similarity 0.80)
- DeNVeR: Deformable Neural Vessel Representations for Unsupervised Video Vessel Segmentation (CVPR 2025 · similarity 0.79)
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.