摘要
针对当前肺结节检测模型精度低、对小结节和近血管结节不敏感的问题,提出了改进的YOLOv3肺结节检测算法。首先,采用Darknet-53 512×512作为主干网络提升对肺内小结节的敏感度;其次,引入Mish激活函数提高模型检测精度;第三,引入K-means++聚类算法,优化锚框,选取更适合肺结节数据集的锚框;第四,将检测框损失函数优化为GIoU,解决了当IoU为0时无法反映预测检测框与真实检测框重合度的情况,同时也避免了当Loss为0时,由于没有梯度回传而无法训练的情况。在LUNA16数据集上的实验结果表明,改进算法经过25 000次迭代后的m AP达到94.89%,比原始YOLOv3算法的m AP提高了4.94%,查准率提高了3.76%,召回率提高了1.15%,对肺内几种常见类型的肺结节都能准确地定位与检测。
Aiming at the problem of low accuracy of current lung nodule detection model and insensitivity to small nodules and proximal vessel nodules,an improved YOLOv3 lung nodule detection algorithm is proposed. Firstly,the Darknet-53 512×512 is used as the backbone network to improve the sensitivity to small nodules in lung. Secondly,the Mish activation function is introduced to improve the accuracy of model detection. Thirdly,the K-means++ clustering algorithm is introduced to optimize the anchor frame and select a more suitable one for lung nodule dataset. Fourthly,the loss function of detection frame is optimized to GIoU,which solves the problem that when IoU is 0,the degree of coincidence between predicted detection frame and real detection frame cannot be reflected,and it also avoids training failure due to no gradient return when Loss is 0. The experimental results of LUNA16 dataset show that the improved algorithm has a m AP of 94. 89% after 25,000 iterations,which is 4. 94% higher than that of original YOLOv3,precision increased by 3. 76%,recall rate increased by 1. 15%.,and several common types of lung nodules can be accurately located and detected.
作者
郭晓敏
黄新
GUO Xiaomin;HUANG Xin(School of Electronic Engineering and Automation,Guilin University of Electronic Science and Technology,Guilin Guangxi 541000,China;Guartgxi Key Laboratory of Automatic Testing Technology and Instrument,Guilin Guangxi 541000,China)
出处
《激光杂志》
CAS
北大核心
2022年第5期207-213,共7页
Laser Journal
基金
国家自然科学基金(No.81873913)
桂林电子科技大学研究生科研创新项目资助(No.2019YCXS101)。