注意力机制与普通深度学习模型对肺结节检测的对比研究  

A comparative study on the detection of pulmonary nodules between the attention mechanism deep learning model and the general deep learning model

作　　者：管乃超[1] 贾俊铖[2] 马文霞[3] 胡春洪[1] GUAN Naichao;JIA Juncheng;MA Wenxia;HU Chunhong(Imaging Center,the First Affiliated Hospital of Soochow University,Suzhou,Jiangsu Province 215006,China;Soochow University School of Computer Science and Technology,Suzhou,Jiangsu Province 215000,China;Department of Quality Management,the First Affiliated Hospital of Soochow University,Suzhou,Jiangsu Province 215006,China)

机构地区：[1]苏州大学附属第一医院影像中心,江苏苏州215006 [2]苏州大学计算机科学与技术学院,江苏苏州215000 [3]苏州大学附属第一医院质量管理处,江苏苏州215006

出　　处：《实用放射学杂志》2021年第10期1605-1609,共5页Journal of Practical Radiology

基　　金：国家重点研发计划项目(2017YFC0114300).

摘　　要：目的对比研究注意力机制与普通深度学习模型对胸部CT肺结节的检测效能.方法分别构建普通深度学习模型和注意力机制深度学习模型,用LUNA16数据集训练模型;回顾性选取295例有肺结节的胸部CT图像测试2个模型,同时1名医师对295例胸部CT阅片诊断,比较3种模式检测肺结节与鉴别良恶性的各项指标.结果对肺结节检测的敏感度、特异度、假阳性率、假阴性率和约登指数,医师为75.35%、86.29%、13.71%、24.65%、0.616;普通深度学习模型为96.0%、57.19%、42.81%、4.0%、0.532;注意力机制深度学习模型为96.8%、83.71%、16.29%、3.2%、0.805.3种模式对恶性肺结节检测敏感度分别为86.5%、84.7%、95.9%.注意力机制深度学习模型检测效能与另外2种模式的差异有统计学意义(P<0.05).结论注意力机制深度学习模型检测肺结节的效能更好,能有效检出恶性肺结节,有助于医师对肺结节的诊断,是目前肺结节影像诊断的较优人工智能(AI)模型.Objective To compare the detection efficiency of the attention mechanism deep learning model and the general deep learning model on chest CT pulmonary nodules.Methods Construct a general deep learning model and an attention mechanism deep learning model,both of which were trained by the LUNA16 dataset.295 cases of chest CT images with pulmonary nodules were reviewed by two models and a medical doctor.The indexes of detecting and differentiating benign and malignant were compared among the three models.ResuIts The sensitivity,specificity,false positive rate,false negative rate,and Yoden index of pulmonary nodule detection were 75.35%,86.29%,13.71%,24.65%,0.616 for doctor;96.0%,57.19%,42.81%,4.0%,0.532 for general deep learning model,and 96.8%,83.71%,16.29%,3.2%,0.805 for attention mechanism deep learning model.The sensitivity of the three models to the detection of malignant pulmonary nodules was 86.5%,84.7%and 95.9%,respectively.The difference of detection efficiency between attention mechanism deep learning model and the other two models was statistically significant(P<0.05).Conclusion The attention mechanism deep learning model has better efficiency in detecting pulmonary nodules and can effectively detect malignant pulmonary nodules.lt is helpful for doctors to diagnose pulmonary nodules,and is a better artificial intelligence(AI)model for imaging diagnosis of pulmonary nodules.

关 键 词：注意力机制深度学习 普通深度学习 肺结节 检测效能 

分 类 号：R814.49[医药卫生—影像医学与核医学] R563[医药卫生—放射医学]