基于改进卷积神经网络的肺部肿瘤检测  

Lung tumor detection based on improved Convolutional neural network

作　　者：王煜君 江宇楠 刘琳岚[1] 张鹏飞 罗坤龙 WANG Yujun;JIANG Yunan;LIU Linlan;ZHANG Pengfei;LUO Kunlong(School of Information Engineering,Nanchang Hangkong University,Nanchang 330063,Jiangxi)

机构地区：[1]南昌航空大学信息工程学院,江西南昌330063

出　　处：《长江信息通信》2024年第1期32-35,共4页Changjiang Information & Communications

摘　　要：肺部肿瘤目前是致死率最高的疾病之一。其初期的形状一般非常微小,且和正常的组织类似,哪怕是经验丰富的医生也无法确保能够精准地确定肿瘤所在位置,因此使用计算机辅助检测是一个不错的选择。在研究深度学习目标检测算法YOLOv5的基础上,针对上述难点,从以下三个方面改进YOLOv5算法来进行肺部肿瘤检测。首先,根据数据集中所有肿瘤大小重新设计了初始检测框的大小;然后,利用主干特征提取网络中的特征图,并新添加了一个检测层,该检测层的特征图只经过了两次下采样,能更好地保留肿瘤的细节信息;最后,选择CBAM注意力机制加入FPN结构,进一步提高模型的检测效果。通过在LUNA16数据集上进行实验,发现改进算法的精确率、召回率和mAP分别达到了96.81%、94.94%和96.6%,比改进前分别提高了15.18%、18.02%和13.46%,且与近三年同类算法相比,也具有较好的检测性能。所以此改进算法能有效地对肺部肿瘤进行检测。At present,Lung tumor is one of the diseases with the highest Case fatality rate rate.Its initial shape is generally very small and similar to normal tissue,and even experienced doctors cannot guarantee the precise location of the tumor.Therefore,using computer-aided detection is agood choice.On the basis of studying the deep learning object detection algorithm YOLOv5,aiming at the above difficulties,YOLOv5 algorithm is improved from the following three aspects to detect Lung tumor.Firstly,the initial detection box size was redesigned based on the size of all tumors in the dataset;Then,the feature maps in the network were extracted using the backbone features,and a new detection layer was added.The feature maps of this detection layer were only downsampled twice,which can better preserve the detailed information of the tumor;Finally,the CBAM attention mechanism was selected to join the FPN structure to further improve the detection performance of the model.Through experiments on the LUNA16 dataset,it was found that the accuracy,recall,and mAP of the improved algorithm reached 96.81%,94.94%,and 96.6%,respectively,which were 15.18%,18.02%,and13.46%higher than before the improvement.Moreover,compared with similar algorithms in the past three years,it also has good detection performance.So this improved algorithm can effectively detect Lung tumor.

关 键 词：肺部肿瘤 目标检测算法 YOLOv5 检测层 注意力机制 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]