Research map / Detection and segmentation / CNN features
Polyper: Boundary Sensitive Polyp 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 Polyper, a two-stage decoder that uses morphological erosion/dilation to split a predicted polyp mask into boundary and interior regions, then refines boundary features with cross-attention to interior features, achieving SOTA polyp segmentation on five datasets.
Problem
Polyp segmentation in colonoscopy struggles with blurred, low-contrast boundaries, especially for small early-stage polyps.
Approach
Use morphological erosion and dilation of the initial segmentation to separate boundary and interior polyp regions, then apply spatial and channel cross-attention so boundary features are refined using interior polyp features (ignoring background tokens to save compute).
Key results
- Polyper achieves 94.82 mDice / 90.36 mIoU on Kvasir and 86.51 mDice / 78.51 mIoU on ETIS, outperforming listed baselines such as SegT and CaraNet.
- Boundary sensitive refinement adds 3.45 mIoU over the initial segmentation with a Swin-T encoder (87.12 to 90.57) in ablation on Kvasir.
- Region separation with 4 erosion iterations improves mIoU by 4.40 over refining the whole foreground mask.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Polyper instead of the strongest listed baseline | Polyp segmentation · Kvasir | mDice: 93.40 -> 94.82 | Helped |
| Use Polyper instead of the strongest listed baseline | Polyp segmentation · ClinicDB | mDice: 94.00 -> 94.45 | Helped |
| Use Polyper instead of the strongest listed baseline | Polyp segmentation · ColonDB | mDice: 80.80 -> 83.72 | Helped |
| Use Polyper instead of the strongest listed baseline | Polyp segmentation · EndoScene | mDice: 90.60 -> 92.43 | Helped |
| Use Polyper instead of the strongest listed baseline | Polyp segmentation · ETIS | mDice: 81.00 -> 86.51 | Helped |
| Add boundary sensitive refinement (ablation) | Polyper with ResNet-50 encoder · Kvasir | mIoU: 82.75 -> 84.97 | Helped |
| Add boundary sensitive refinement (ablation) | Polyper with MiT-B1 encoder · Kvasir | mIoU: 87.74 -> 89.19 | Helped |
| Add boundary sensitive refinement (ablation) | Polyper with Swin-T encoder · Kvasir | mIoU: 85.16 -> 90.36 | Helped |
| Add region separation instead of refining the full foreground (ablation) | Polyper · Kvasir | mIoU: 86.55 -> 90.57 | Helped |
| Use 4 erosion iterations rather than 5 (ablation) | Polyper region separation · Kvasir | mIoU: 89.66 -> 90.57 | Helped |
| Add spatial attention (ablation) | Polyper BSA with no attention branches · Kvasir | mIoU: 87.65 -> 89.45 | Helped |
| Add channel attention (ablation) | Polyper BSA with no attention branches · Kvasir | mIoU: 87.65 -> 89.16 | Helped |
Surprising findings
- Non-Local and Hamburger feature aggregation underperform the proposed aggregation on medical data, which the authors attribute to limited medical training data.
Limitations
- Assumes polyp localization is accurate; struggles with false positives and false negatives.
- Uses a fixed-width boundary definition that may not fit diverse polyp features.
Experimental setup
Datasets: Kvasir-SEG CVC-ClinicDB
Benchmarks: Kvasir-SEG CVC-ClinicDB CVC-ColonDB EndoScene ETIS
Metrics: mDice mIoU
Baselines: U-Net UNet++ SFA ACSNet PraNet SANet TransFuse TransUNet HarDNet-MSEG DS-TransUNet SwinE-Net Polyp-PVT
Builds on: Swin Transformer CaraNet Non-Local Hamburger
Closest papers by meaning
- Boost the Inference with Co-training: A Depth-guided Mutual Learning Framework for Semi-supervised Medical Polyp Segmentation (CVPR 2025 · similarity 0.87)
- MonoBox: Tightness-Free Box-Supervised Polyp Segmentation Using Monotonicity Constraint (AAAI 2025 · similarity 0.84)
- An Embedding-Unleashing Video Polyp Segmentation Framework via Region Linking and Scale Alignment (AAAI 2024 · similarity 0.84)
- Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.83)
- Domain Generalized Medical Landmark Detection via Robust Boundary-Aware Pre-Training (AAAI 2025 · similarity 0.82)
- ConDSeg: A General Medical Image Segmentation Framework via Contrast-Driven Feature Enhancement (AAAI 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.