一种融合肺炎影像特征的肺结核病灶检测迁移网络  

A transfer network for pulmonary tuberculosis lesions detection on the fusion of image features of pneumonia

作　　者：安乐 彭柯鑫[1] 杨兴 黄盼 魏彪[2,3] 冯鹏[2,3] AN Le;PENG Kexin;YANG Xing;HUANG Pan;WEI Biao;FENG Peng(College of Computer Science and Cyber Security,Chengdu University of Technology,Chengdu 610059,China;The Key Lab of Optoelectronic Technology and Systems,Ministry of Education,Chongqing University,Chongqing 400044,China;Engineering Research Center of Industrial CT Nondestructive Testing,Ministry of Education,Chongqing University,Chongqing 400044,China)

机构地区：[1]成都理工大学计算机与网络安全学院图像信息处理研究室,四川成都610059 [2]重庆大学光电工程学院光电技术及系统教育部重点实验室,重庆400044 [3]重庆大学光电工程学院工业ICT无损检测教育部工程研究中心,重庆400044

出　　处：《光学技术》2023年第1期120-128,共9页Optical Technique

基　　金：科技部重点研发计划(2019YFC0605203);重庆市研究生研究科研创新项目(CYS22109、CYS22111);重庆市科委基础研究与前沿探索项目(cstc2020jcyj-msxmX0553);重庆市科委技术创新与应用发展专项(cstc2021jscx-gksbX0056)。

摘　　要：基于胸部X光透射图像(DR)的肺部病灶识别与疾病诊断是临床的常规操作。对于肺结核患者而言,其DR图像中的病灶区域与背景相融性高,目标弥散严重且边缘形态极不规则,严重干扰诊断的准确性。针对上述问题,提出了一种融合肺炎影像学特征的肺结核病灶区域检测网络(TDT-Net),利用肺结核和新冠肺炎同为呼吸道疾病且在DR图像上具有相似表征的特点,借助大量肺炎DR数据,通过迁移学习引入强相关特征以提高肺结核病灶的检测精度。TDT-Net结合Transformer和扩张残差技术,提出上下文感知增强模块,以强化迁移模型对全局信息的建模能力;利用特征细化模块减少迁移过程中引入的冗余信息,凸显强相关特征的表示。实验结果表明,在TBX11K数据集上,所提检测方法的平均准确度(AP)达到87.5%,召回率(Recall)达到80.7%,相较于YOLOV5和RetinaNet等网络有效提升了肺结核病灶的检测精度,实现了更加准确的肺结核病灶定位和分类。Lung lesion detection and disease diagnosis based on chest X-ray images(DR) is a routine clinical operation. For pulmonary tuberculosis patients, the lesion area of tuberculosis in the DR image is highly compatible with the background, the target diffusion is serious, and the edge shape is extremely irregular, which seriously interferes with the accuracy of diagnosis. To solve the above problems, a Tuberculosis Deep Transfer Net(TDT-Net)integrating the imaging characteristics of pneumonia is proposed. Using the characteristics that tuberculosis and COVID-19 are respiratory infectious diseases and have similar imaging manifestations on DR images,with the help of a large number of pneumonia DR images,strong correlation features are introduced through transfer learning to improve the detection accuracy of tuberculosis lesions.TDT-Net combines transformer and extended residual technology,and proposes a context-aware enhancement module to strengthen the modeling ability of the migration model for global context information;The feature refinement module is used to reduce the redundant information introduced in the transfer process and highlight the representation of strongly related features.The experimental results show that on the TBK11 kdataset,the Average Precision(AP)of the proposed detection method reaches 87.5%,and the Recall reaches 80.7%.Compared with the networks such as YOLOV5 and RetinaNet,the detection accuracy of tuberculosis lesions is effectively improved,and more accurate localization and classification of tuberculosis lesions are achieved.

关 键 词：X射线成像 肺结核病灶 迁移学习 目标检测 肺炎特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]