基于U-NET深度学习算法检测与量化胸部CT图像的气胸  

Pneumothorax detection and quantification in chest computed tomography images using a U-NET deep learning segmentation algorithm

作　　者：王舒皓 叶丽芳 杨茗茗 孟以爽 李跃华 WANG Shu-hao;YE Li-fang;YANG Ming-ming(Department of Radiology,the Sixth People's Hospital affiliated to Shanghai Jiaotong University School of Medicine,Shanghai 200233,China)

机构地区：[1]上海交通大学医学院附属第六人民医院放射科,上海200233 [2]飞利浦中国投资有限公司,上海200072

出　　处：《放射学实践》2023年第3期290-294,共5页Radiologic Practice

摘　　要：目的:采用气胸患者的CT图像开发了一种使用U-NET架构的深度学习分割算法,并对其性能进行了评估。方法:回顾性分析2018年-2019年的急诊胸部CT图像,先由一名放射科医生进行注释,然后由另一名资深放射科医生修改和审查标注的气胸内容,作为金标准,并使用五折交叉验证方法进行深度学习算法的训练和测试。在像素级通过戴斯系数、召回率和符合率来评估分割精度,并评估了气胸定量的体积误差;在区域级评估每个患者的气胸区域敏感性和假阳性区域数量。结果:共有200例气胸患者入组,平均戴斯系数、召回率和符合率分别为0.789、0.794和0.820。对气胸总量大于300 mL的患者,平均戴斯系数、召回率和符合率都可以达到0.89以上。对气胸总量大于100 mL的患者,气胸定量的相对误差小于10%。对体积大于30 mL的气胸区域,区域敏感性可达100%。假阳性区域平均体积为2.3 mL(1.55~3.66 mL),平均每个病例2.8(2.06~4.23)个假阳性区域。结论:U-NET深度学习分割算法在像素和区域两个层面上都表现出可接受的性能,这表明在临床实践中可以发挥潜在的辅助作用,以减轻急诊医务人员的工作量。Objective:To construct a deep learning segmentation algorithm using U-NET architecture based on CT images of pneumothorax patients,and evaluate its performance.Methods:The emergency chest CT images from 2018-2019 were retrospective collected and annotated by a radiologist,followed by another senior radiologist to revise and review the annotated pneumothorax content as a gold standard.A deep learning algorithm was trained and tested using a five-fold cross-validation method.Segmentation accuracy was evaluated at the pixel level by Dice coefficients,recall and compliance,and volumetric error in pneumothorax quantification was assessed.The sensitivity of pneumothorax regions and the number of false positive regions were evaluated at the region level for each patient.Results:A total of 200 pneumothorax patients were enrolled,and the mean Dice coefficient,recall rate and compliance rate were 0.789,0.794,and 0.820,respectively.For patients with total pneumothorax greater than 300mL,the average Dice coefficient,recall and compliance rate can reach more than 0.89.For patients with a total pneumothorax volume greater than 100mL,the relative error of pneumothorax quantification was less than 10%.For the pneumothorax region with volume greater than 30mL,the regional sensitivity could reach 100%.The average volume of false-positive regions was 2.3mL(1.55~3.66mL),and the average number of false-positive regions was 2.8(2.06~4.23)per case.Conclusion:The U-NET deep learning segmentation algorithm showed acceptable performances at both pixel and region levels,which can suggest as a potential tool in assisting emergency medical personnel to reduce the workload.

关 键 词：体层摄影术 X线计算机 人工智能 深度学习 气胸 

分 类 号：R814.42[医药卫生—影像医学与核医学] R-05[医药卫生—放射医学] R561.4[医药卫生—临床医学]