基于形状特征引导的息肉分割网络
Shape Features Guided Network for Polyp Segmentation
吴英豪 1杜晓刚 1雷涛 1张学军 2王营博1
作者信息
- 1. 陕西科技大学 人工智能联合实验室,陕西 西安 710021;陕西科技大学 电子信息与人工智能学院,陕西 西安 710021
- 2. 兰州交通大学 电子与信息工程学院,甘肃 兰州 730070
- 折叠
摘要
由于息肉的形状、大小不一,且息肉与周围粘膜的对比度较低,导致息肉分割是一个极具挑战性的任务.为了提升在不同形状息肉和小尺寸息肉处的分割精度,提出了一个基于形状特征引导的息肉分割网络(SGNet).SGNet主要包括两点贡献:首先,通过引入ConvNext设计了细节敏感编码器,使网络在提取全局信息的同时也能提取到对息肉分割至关重要的细节信息.其次,设计了形状引导解码器,该解码器能有效适应不同形状的息肉,提取到丰富的息肉形状特征和多尺度特征,从而有效提升网络在不同形状息肉和小尺寸息肉的分割精度.在三个公开的息肉分割数据集ETIS、Kvasir和CVC-ClinicDB上进行大量实验表明,SGNet能够精确地分割不同形状的息肉和小尺寸息肉,在分割精度上超过了近年来主流的息肉分割网络.
Abstract
Polyp segmentation is a challenging task due to the variable size and shape of polyps and the low contrast between polyps and surrounding mucosa.To improve the segmentation accuracy at polyps with different shapes and tiny-polyps,we propose a shape-guided polyp segmentation network namely SGNet.SGNet mainly includes two contributions:Firstly,we design a detail-sensitive encoder by in-troducing ConvNext,so that the network can extract the details that are crucial to polyp segmentation while extracting global information.Secondly,we design a shape-guided decoder,which can not only effectively adapt to polyps with different shapes,but also extract rich polyp shape features and multi-scale features,thus effectively improving the segmentation accuracy in polyps with different shapes and tiny-polyps.Extensive experiments on three public polyp segmentation datasets(ETIS,Kvasir and CVC-ClinicDB)show that SGNet can accurately segment polyps with different shapes and tiny-polyps,and is superior to the popular polyp segmentation networks in terms of segmentation accuracy in recent years.
关键词
深度学习/卷积神经网络/医学图像分割/息肉分割/形状特征Key words
deep learning/convolutional neural network/medical image segmentation/polyp segmentation/shape feature引用本文复制引用
基金项目
国家自然科学基金(61861024)
国家自然科学基金(62271296)
国家自然科学基金(62201334)
出版年
2024
国家科技期刊平台
NSTL
万方数据