MICCAI 2026 logo Southeast University logo Mohamed bin Zayed University of Artificial Intelligence logo

CMSA-Net (MICCAI 2026)

Causal Multi-scale Aggregation with Adaptive Multi-source Reference for Video Polyp Segmentation

Tong Wang1,2 Yaolei Qi1 Siwen Wang2 Imran Razzak2 Guanyu Yang1 Yutong Xie2

1Southeast University    2Mohamed bin Zayed University of Artificial Intelligence

Paper Code Poster Slides Weights Results
CMSA-Net framework diagram showing input colonoscopy video, adaptive multi-source reference, causal multi-scale aggregation, segmentation decoder, and current-frame prediction.

CMSA-Net addresses weak semantic discrimination, motion blur, and scale variation through causal temporal modeling and adaptive reference selection.

  • CMACausal multi-scale aggregation gathers historical evidence while preserving temporal order.
  • DMRDynamic multi-source reference selection chooses reliable semantic and local references.
  • VPSStable video polyp segmentation under appearance, motion, and scale changes.

Presentation

English and Chinese video explanations

Conference-style walkthroughs of the motivation, architecture, experiments, and key conclusions.

English CMSA-Net MICCAI 2026 presentation
中文 CMSA-Net 中文讲解

Visual results

Cleaner masks across hard video frames.

Qualitative comparisons show that CMSA-Net maintains more complete polyp masks than baseline predictions under ambiguity, blur, and appearance variation.

Qualitative CMSA-Net results compared with ground truth and baselines across multiple colonoscopy frames.
Qualitative comparison on video polyp segmentation samples.

Citation

Cite CMSA-Net

@article{wang2026cmsanet,
  title={CMSA-Net: Causal Multi-scale Aggregation with Adaptive Multi-source Reference for Video Polyp Segmentation},
  author={Tong Wang and Yaolei Qi and Siwen Wang and Imran Razzak and Guanyu Yang and Yutong Xie},
  journal={arXiv preprint arXiv:2602.22821},
  year={2026}
}