融合非局部注意力机制的旋转病灶检测算法  

Rotating Lesion Detection Algorithm Fused with Non-localAttention Mechanism

作　　者：陈倩 詹峰 王波 郭一娜[1] CHEN Qian;ZHAN Feng;WANG Bo;GUO Yi-na(College of Electronic Information Engineering,Taiyuan University of Science and Technology,Taiyuan 030024,China;College of Engineering,Fujian Jiangxia University,Fuzhou 350108,China)

机构地区：[1]太原科技大学电子信息工程学院,太原030024 [2]福建江夏学院工程学院,福州350108

出　　处：《太原科技大学学报》2025年第1期14-19,共6页Journal of Taiyuan University of Science and Technology

基　　金：国家自然科学基金(62271341);国家留基委地区合作与高层次人才培养项目([2020]1417);山西省自然科学优秀青年基金(201901D211313);山西省回国留学人员科研资助项目(HGKY2019080);福建省自然科学基金(2020J01937);山西省研究生优秀创新项目计划(2022Y689)。

摘　　要：病灶检测旨在对医学图像的病灶部位进行检测、定位和识别。针对目前病灶检测算法中存在的小目标漏检、病灶方向单一和目标混叠问题,提出了一种融合非局部注意力机制的旋转病灶检测算法,实现病灶类型和位置的精确检测。算法中主干网络中引入的非局部注意力机制,获取特征图中像素之间的全局依赖,提升小病灶的检测精度;设计角度敏感网络,通过旋转角度的设置,识别多个方向的旋转病灶,弥补水平边界框在多角度目标检测中的不足;构建RNMS (Rotation Non-maximum Suppression)和MRNMS(Multi-type Rotation Non-maximum Suppression)算法,通过排序、置信度对比、计算IoU(Intersection over Union)等操作,有效地解决了单发性病灶检测过程中的同类别目标混叠及不同类别目标混叠现象。在Rotated PET-CT-Dx数据集的结果显示,mAP达到87.3%.Lesion detection aims to detect,locate and identify the lesion in medical images.In view of the problems of missed detection of small lesions,different shapes of lesions and aliasing of targets in the current lesion detection algorithm,in this paper,we propose a rotating lesion detection algorithm fused with non-local attention mechanism to achieve accurate detection of lesion type and location.The improved backbone network can identify small lesions,non-local attention mechanism added in each block,global dependencies among pixels in the feature map are obtained,improved detection accuracy for small lesions.Angle sensitive network by setting the rotation angle,can accurately identify rotating lesions in multiple directions,which make up the deficiency of horizontal bounding boxes in multi-angle object detection.Aiming at the problem of target aliasing in same type or different types,RNMS(Rotated Non-Maximum Suppression)and MRNMS(Multi-type Rotated Non-Maximum Suppression)are proposed to reduce the target aliasing in single lesion detection.Experimental results show that the proposed algorithm outperforms state-of-the-art algorithms on Rotated PET-CT-Dx dataset,mAP can reach 87.3%.

关 键 词：旋转病灶检测 非局部注意力机制 角度敏感网络 RNMS MRNMS 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]