文献类型：期刊文献

中文题名：面向内窥影像息肉区域感知的分层多尺度特征归并融合分析方法

英文题名：A Hierarchical Multi-Scale Feature Aggregation and Fusion Method for Polyp Region Perception in Endoscopy Image

作者：胡珂立[1,2,3];黄杰[1];孙源泽[1];何为[4];祝汉灿[1,2];王臣[1,2];赵利平[1];胡剑浩[3]

机构：[1]绍兴文理学院计算机科学与工程系,浙江绍兴312000;[2]绍兴文理学院人工智能研究院,浙江绍兴312000;[3]绍兴文理学院附属医院,浙江绍兴312000;[4]中国科学院上海微系统与信息技术研究所,上海200050

年份：2025

卷号：38

期号：8

起止页码：1448

中文期刊名：传感技术学报

外文期刊名：Chinese Journal of Sensors and Actuators

收录：北大核心2023、、北大核心

基金：浙江省自然科学基金重点项目(LZ24F020006);国家自然科学基金资助项目(62271321,61603258)。

语种：中文

中文关键词：内窥镜检查;区域感知;结直肠癌;息肉;提取

外文关键词：endoscopy;regional perception;colorectal cancer;polyp;extraction

中文摘要：结肠镜是肠道内窥感知的重要手段,是结直肠癌筛查的金标准,及时发现并进行必要摘除是降低结直肠癌发病率的重要手段。针对息肉区域难以准确提取的问题,本文提出了一种基于分层多尺度特征归并融合的内窥影像息肉区域感知提取网络,以帮助医生完整发现并摘除相应组织区域。首先,该网络引入非均衡信息聚合模块,通过不同尺度卷积和全局平均池化特征进行相邻层特征融合。接着,通过多尺度信息聚合模块,结合多卷积核提取特征并基于通道注意力机制强化不同感受野特征融合提取。最后,基于两类模块逐级串联归并,强化各层多尺度特征归并融合,增强网络对模糊边界和小病灶区域的识别能力。在5个数据集上,经比较分析,本文提出的网络总体表现最优。本文提出的网络在CVC-ColonDB和ETIS两个数据集中表现出较大优势,在平均Dice指标上相比CFANet最高取得了7.7%的提升,验证了算法有效性。

外文摘要：Colonoscopy is a critical tool for intestinal endoscopic examination and is regarded as the gold standard for colorectal cancer screening.Early detection and timely removal of lesions are vital in reducing the incidence of colorectal cancer.To address the issue of accurately extracting polyp regions,a hierarchical multi-scale feature merging and fusion network for polyp region perception and extrac-tion in endoscopic images is introduced,which is designed to assist physicians in accurately identifying and removing relevant tissue are-as.First,a non-balanced information aggregation module is proposed,which integrates adjacent-layer features through multi-scale convo-lutions and global average pooling.Second,a multi-scale information aggregation module is introduced,leveraging multi-kernel convolu-tional feature extraction combined with a channel attention mechanism to enhance the fusion of features from different receptive fields.Finally,the two modules are sequentially connected to progressively merge multi-scale features across layers,strengthening the networks ability to detect regions with blurred boundaries and small lesions.Comprehensive evaluations across five datasets highlight the superior overall performance of the proposed network.Notably,on the CVC-ColonDB and ETIS datasets,the proposed network shows significant advantages,achieving up to a 7.7%improvement in the average Dice coefficient compared to CFANet,and this validates the effective-ness of the proposed method.