Research map / Image generation / Fast diffusion sampling
MedFlowSeg: Flow Matching for Medical Image Segmentation with Frequency-Aware Attention
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 MedFlowSeg, a conditional flow matching model for medical image segmentation with a dual-branch frequency-aware conditioning design (DB-SA and FA-Attention), evaluated on ACDC, BraTS-2021, REFUGE2, GlaS and CAMUS.
Problem
Diffusion-based medical image segmentation requires many iterative sampling steps and struggles to integrate conditional image features with noisy intermediate states, limiting efficiency and cross-modality generalization.
Approach
Formulates segmentation as conditional flow matching with a UNet flow network and a parallel UNet condition network; injects multi-frequency (low/high-pass) structural priors into the shallow flow encoder (DB-SA) and models spatial and spectral (FFT) discrepancies between flow and condition features at the bottleneck with time-modulated fusion (FA-Attention).
Key results
- On ACDC, MedFlowSeg reaches 93.57 average Dice and 5.27 HD95, versus 92.68 Dice for the strongest baseline MedSegDiff-V2.
- On BraTS-2021, MedFlowSeg reports 91.2 Dice and 6.76 HD95, versus 90.8 Dice for MedSegDiff-V2.
- Ablation on three datasets shows DB-SA and FA-Attention each add gains over a baseline (e.g., Brain Tumor Dice 87.6 to 91.2 with full model).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DB-SA structural condition (multi-frequency priors into first encoder stage) (ablation) | flow network backbone · Ablation set (brain tumor, optic cup, histology gland) | Dice: +1.4 brain, +2.2 optic cup, +3.2 gland | Helped |
| Add standard spatial attention (SA) instead of DB-SA (ablation) | flow network backbone · Ablation set (brain tumor, optic cup, histology gland) | Dice: +0.9 brain, +0.8 optic cup, -0.2 gland | Helped |
| Add FA-Attention semantic conditioning (without Neural Modulator) on top of DB-SA (ablation) | flow network with DB-SA · Ablation set (brain tumor, optic cup, histology gland) | Dice: +1.7 brain, +1.5 optic cup, +1.2 gland | Helped |
| Add Neural Modulator on top of FA-Attention (ablation) | MedFlowSeg · Ablation set (brain tumor, optic cup, histology gland) | Dice: +0.5 brain, +0.7 optic cup, +1.3 gland | Helped |
| Combine DB-SA and FA-Attention (full MedFlowSeg) vs. MedSegDiff-V2 baseline | MedFlowSeg · ACDC | Dice: 93.57 vs 92.68 (MedSegDiff-V2) | Helped |
| Flow matching ODE sampling with 50 steps instead of 100-step diffusion sampling | MedFlowSeg · REFUGE2-Cup / efficiency | Gflops: 789 vs 983 (MedSegDiff-V2) | Helped |
| Use of MedFlowSeg on Disc segmentation (REFUGE2-Disc) | MedFlowSeg · REFUGE2-Disc | Dice: 95.8 vs 96.7 (MedSegDiff-V2) | Hurt |
| Replace MSE with L1 velocity loss | MedFlowSeg · ACDC | Dice | No clear effect |
Surprising findings
- Transformer-based Swin-UNet does not consistently beat nnUNet, and early diffusion methods (EnsemDiff, SegDiff) do not reliably surpass deterministic baselines.
Limitations
- Gains over MedSegDiff-V2 are modest on several datasets (e.g., REFUGE2-Cup Dice 88.6 vs 87.9, but IoU lower than MedSegDiff-V2 on Cup), and results are single-run without reported variance.
- Ensemble of 10 samples fused by STAPLE is still required, so inference is not single-pass.
- Parameter count (111M) is larger than several baselines such as FlowSDF (42M).
Experimental setup
Datasets: ACDC BraTS-2021 REFUGE2 GlaS CAMUS
Benchmarks: ACDC BraTS-2021 REFUGE2-Disc REFUGE2-Cup GlaS CAMUS
Metrics: Dice IoU HD95 GED CI Gflops
Baselines: UNet nnUNet TransUNet SwinUNet MT-UNet EnsemDiff SegDiff MedSegDiff MedSegDiff-V2 FlowSDF TransBTS SwinBTS
Builds on: conditional flow matching MedSegDiff UNet STAPLE
Closest papers by meaning
- Rethinking Vector Field Learning for Generative Segmentation (NeurIPS 2026 · similarity 0.86)
- STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation (NeurIPS 2026 · similarity 0.86)
- FAMNet: Frequency-aware Matching Network for Cross-domain Few-shot Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- Structured Coupling for Flow Matching (NeurIPS 2026 · similarity 0.85)
- MedSegDiff-V2: Diffusion-Based Medical Image Segmentation with Transformer (AAAI 2024 · similarity 0.85)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.85)
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.