Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PolypFlow: Reinforcing Polyp Segmentation with Flow-Driven Dynamics

Pu Wang, Huaizhi Ma, Zhihua Zhang, Zhuoran Zheng

TL;DR

PolypFlow tackles the challenge of segmenting polyps with irregular shapes and boundaries by introducing a flow-matching refinement that solves the ODE $d\phi_t/dt = v_t(\phi_t)$ to progressively align coarse predictions with ground truth. The vector field $v$ blends self-attention-driven weighting with frequency-domain priors (DCT) and is integrated with U-Net features to produce an interpretable, trajectory-based refinement over $N=10$ steps, guided by a loss that combines weighted IoU and BCE terms. This approach yields boundary-aware robustness and clear visualization of the refinement process, and achieves state-of-the-art performance across five benchmarks with strong generalization to unseen datasets. The results indicate substantial gains from incorporating flow dynamics and frequency-aware features, suggesting practical impact for reliable, edge-friendly polyp segmentation in diverse clinical conditions.

Abstract

Accurate polyp segmentation remains challenging due to irregular lesion morphologies, ambiguous boundaries, and heterogeneous imaging conditions. While U-Net variants excel at local feature fusion, they often lack explicit mechanisms to model the dynamic evolution of segmentation confidence under uncertainty. Inspired by the interpretable nature of flow-based models, we present \textbf{PolypFLow}, a flow-matching enhanced architecture that injects physics-inspired optimization dynamics into segmentation refinement. Unlike conventional cascaded networks, our framework solves an ordinary differential equation (ODE) to progressively align coarse initial predictions with ground truth masks through learned velocity fields. This trajectory-based refinement offers two key advantages: 1) Interpretable Optimization: Intermediate flow steps visualize how the model corrects under-segmented regions and sharpens boundaries at each ODE-solver iteration, demystifying the ``black-box" refinement process; 2) Boundary-Aware Robustness: The flow dynamics explicitly model gradient directions along polyp edges, enhancing resilience to low-contrast regions and motion artifacts. Numerous experimental results show that PolypFLow achieves a state-of-the-art while maintaining consistent performance in different lighting scenarios.

PolypFlow: Reinforcing Polyp Segmentation with Flow-Driven Dynamics

TL;DR

PolypFlow tackles the challenge of segmenting polyps with irregular shapes and boundaries by introducing a flow-matching refinement that solves the ODE to progressively align coarse predictions with ground truth. The vector field blends self-attention-driven weighting with frequency-domain priors (DCT) and is integrated with U-Net features to produce an interpretable, trajectory-based refinement over steps, guided by a loss that combines weighted IoU and BCE terms. This approach yields boundary-aware robustness and clear visualization of the refinement process, and achieves state-of-the-art performance across five benchmarks with strong generalization to unseen datasets. The results indicate substantial gains from incorporating flow dynamics and frequency-aware features, suggesting practical impact for reliable, edge-friendly polyp segmentation in diverse clinical conditions.

Abstract

Accurate polyp segmentation remains challenging due to irregular lesion morphologies, ambiguous boundaries, and heterogeneous imaging conditions. While U-Net variants excel at local feature fusion, they often lack explicit mechanisms to model the dynamic evolution of segmentation confidence under uncertainty. Inspired by the interpretable nature of flow-based models, we present \textbf{PolypFLow}, a flow-matching enhanced architecture that injects physics-inspired optimization dynamics into segmentation refinement. Unlike conventional cascaded networks, our framework solves an ordinary differential equation (ODE) to progressively align coarse initial predictions with ground truth masks through learned velocity fields. This trajectory-based refinement offers two key advantages: 1) Interpretable Optimization: Intermediate flow steps visualize how the model corrects under-segmented regions and sharpens boundaries at each ODE-solver iteration, demystifying the ``black-box" refinement process; 2) Boundary-Aware Robustness: The flow dynamics explicitly model gradient directions along polyp edges, enhancing resilience to low-contrast regions and motion artifacts. Numerous experimental results show that PolypFLow achieves a state-of-the-art while maintaining consistent performance in different lighting scenarios.

Paper Structure

This paper contains 11 sections, 8 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Preliminaries
  5. U-Net: Structure and Output
  6. Flow Matching for Polyp Segmentation
  7. Experiment
  8. Datasets and Training Settings.
  9. Performance Comparisons
  10. Ablation Study and Discussion
  11. Conclusion

Figures (3)

  • Figure 1: Overview of PolypFlow. (a) The overall training and inference process of PolypFlow based on flow matching. (b) The ordinary differential equation trajectory begins from a data-dependent prior distribution. (c) Vector field based on Self-attention and DCT. It is worth noting that our vector field is a core step in feature extraction, focusing on local features (convolution), global features (Self-attention), and the frequency domain (DCT).
  • Figure 2: Visualize the results at each step.
  • Figure 3: visual results of different methods.