摘要
在结肠镜检查中,自动分割息肉是开发计算机辅助结肠镜检测和诊断系统的先决条件。息肉分割是一项非常具有挑战性的任务,因为息肉与周边组织具有很大的相似性以及息肉的大小形状变化很大。针对息肉与周围组织相似以及息肉多变的问题,提出基于全局线索定位和多视图特征融合的息肉分割方法。设计全局线索定位模块将全局定位信息传播到每个级别的特征图中,以显式的方式使每个级别的特征图都获得伪装特性息肉的位置信息;设计自我多视图特征融合模块,通过不同视图下的特征捕获不同视图之间的层次特征,更好地适应不同情况下的息肉分割场景。提出方法在5个数据集上的得分比对比算法提高的百分点分别为1.2、3.3、1.8、8.5和3.7,证明提出方法在学习能力和泛化能力上都达到了预期的效果。
In colonoscopy,automatic polyp segmentation is a prerequisite for developing computer-aided colonoscopy detection and diagnosis systems.Polyp segmentation is a highly challenging task due to the high similarity between polyps and surrounding tissues,as well as the significant variability in polyp size and shape.To address these challenges,this paper proposes a polyp segmentation method based on global clue localization and multi-view feature fusion.A global clue localization module is designed to propagate global positional information to each level of feature maps,explicitly embedding positional knowledge of camouflaged polyps into the features.Additionally,a self multi-view feature fusion module is developed to capture hierarchical features across different views,enabling better adaptability to various polyp segmentation scenarios.The proposed method outperforms PraNet by margins of 1.2,3.3,1.8,8.5,and 3.7 percentage points on five datasets,demonstrating its effectiveness in terms of learning and generalization capabilities.
作者
简丽琼
李春生
陈志莉
车进
白雪冰
高翔
JIAN Liqiong;LI Chunsheng;CHEN Zhili;CHE Jin;BAI Xuebing;GAO Xiang(Ningxia Blood Center,Yinchuan 750001,P.R.China;Health Commission Information Center of Ningxia Hui Autonomous Region,Yinchuan 750001,P.R.China;School of Electronic and Electrical Engineering,Ningxia University,Yinchuan 750021,P.R.China;School of Advanced Interdisciplinary Studies,Ningxia University,Zhongwei 755000,P.R.China)
出处
《重庆邮电大学学报（自然科学版）》
CSCD
北大核心
2024年第6期1205-1214,共10页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
国家自然科学基金项目(61861037)
宁夏自然科学基金项目(2022AAC03704)。
关键词
结肠镜
息肉分割
全局线索定位
自我多视图特征融合
伪装
colonoscopy
polyp segmentation
global cue location
self-multiview feature fusion
camouflage
