人工智能
计算机科学
分割
图像分割
模式识别(心理学)
代表(政治)
计算机视觉
特征提取
特征(语言学)
尺度空间分割
特征学习
深度学习
医学影像学
图像处理
基于分割的对象分类
可视化
目标检测
作者
Zongrui Yin,Xuejing Qiao,Jianying Tian,Chenglei Sun,Jiantao Peng,Jianchao Cui,Jindong Xu
标识
DOI:10.1109/jbhi.2026.3747751
摘要
Accurate polyp segmentation is important for computer-aided colorectal cancer screening, yet it remains challenging because of ambiguous boundaries, irregular morphologies, small lesion regions, and the visual similarity between polyps and surrounding mucosal tissues. Existing methods mainly enhance local details or global semantics at the whole-image level, while regional observations are often processed independently or reconstructed without explicitly considering repeated spatial coverage. To address this issue, we propose ORR-Net, an overlap-aware regional representation learning framework for polyp segmentation. ORR-Net constructs dense overlapping regional observations together with a downsampled global observation, allowing the same image location to be represented under different neighboring contexts. A CNN-Transformer dual-branch encoder extracts full-image convolutional features and hierarchical regional-global representations. During decoding, Overlapping Regional Context Attention (ORCA) introduces global contextual information into the regional representations. Overlap-Aware Average Reconstruction (OAAR) further aggregates repeatedly covered positions to recover a continuous full-image representation. A coarse structural prior and multi-level adaptive supervision are employed to support final prediction and intermediate optimization. Experiments on five public polyp segmentation benchmarks demonstrate that ORR-Net achieves competitive or leading performance in segmentation accuracy and boundary quality. The method also shows consistent improvements on unseen cross-dataset test sets. Ablation studies verify the effectiveness of the overlapping regional observation strategy, regional-global interaction, overlap-aware reconstruction, and the associated training designs.
科研通智能强力驱动
Strongly Powered by AbleSci AI