摘要
针对目前相关研究存在只检测特定类型病灶及传统目标检测框架存在固有局限等问题,提出一种针对多类型小目标病灶的检测网络。基于Mask R-CNN进行改进,以融合注意力机制的卷积网络为主干网络,通过计算有效感受野与真实标注框的匹配程度进行正负样本的标签分配,级联多个检测器循环修正回归框。将提出的方法在DeepLesion数据集和外部验证集上进行实验,其结果表明,该模型可以快速准确地对多类型小目标病灶进行检测。
Aiming at the problems that the current researches only detect specific types of lesions and traditional target detection frameworks have inherent limitations,a detection network for multiple types of small target lesions was proposed.This model was improved based on Mask R-CNN,with a convolutional network fused with attention mechanism as the backbone network.By calculating the matching degree between the effective receptive field and the real annotation box,the label allocation of positive and negative samples was performed.Multiple detectors were cascaded to cyclically correct the regression box.The proposed method was tested on the DeepLesion dataset and external validation set.It is indicated that the model can quickly and accurately detect multiple types of small lesions.
作者
张茜
李若宣
郑冰洁
ZHANG Qian;LI Ruo-xuan;ZHENG Bing-jie(School of Artificial Intelligence,Zhongyuan University of Technology,Zhengzhou 450007,China;School of Computer Science,Zhongyuan University of Technology,Zhengzhou 450007,China;Department of Radiology,Affiliated Cancer Hospital of Zhengzhou University,Zhengzhou 450008,China)
出处
《计算机工程与设计》
北大核心
2025年第4期1227-1233,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(62206252,82202270)
河南省自然科学基金青年科学基金项目(252300420995)
河南省高等学校重点科研项目计划基金项目(24B520048)
中原工学院研究生科研创新基金项目(YKY2023ZK40)
中原工学院优势学科实力提升计划基金项目(SD202230)。
关键词
深度学习
感受野
注意力机制
级联结构
多类型
小目标
病灶检测
deep learning
receptive field
attention mechanism
cascade structure
multiple types
small target
lesion detection
