基于人工智能的轻型输入性新型冠状病毒肺炎患者胸部CT的影像学分析

Imaging analysis of chest CT of mild imported corona virus disease 2019 patients based on artificial intelligence

目的 初步探讨轻型输入性新型冠状病毒肺炎(简称新冠肺炎)患者的胸部CT影像学特点.方法 应用人工智能(AI)深度学习方法,对深圳市部分地区52例轻型输入性新冠肺炎患者的影像学特点及临床资料进行回顾性分析,同时由2名资深影像专业医师独立对每个图像结果进行分析及确认;总结人工阅片和AI胸部CT分析的特征,比较人工阅片与AI胸部CT分析特点的差异.结果 相比人工阅片,AI胸部CT分析能自动检测并标记感染区域,同时对感染区域进行量化分析.人工阅片结果显示,所有患者胸部CT均可见异常征象,75.00%(39/52)病变呈双肺散在分布,25.00%(13/52)病变呈单一肺叶或肺段分布,67.31%(35/52)病变多累及双肺下叶及胸膜下;73.07%(38/52)CT表现为磨玻璃密度影,38.46%(20/52)为斑片状渗出影,28.84%(15/52)为实变密度影,与AI胸部CT图像分析结果均一致.基于AI的胸部CT图像分析还显示,新冠肺炎病灶受累区域大多数为双肺下叶,感染体积〔右肺(mm3):54857.13±23978.90比46663.24±16898.62,左肺(mm3):34933.61±16400.66比2348.57±1241.08〕和感染比例〔右肺(%):7.19±5.40比6.15±2.13,左肺(%):3.24±1.62比0.19±0.11〕均高于上叶;受累肺段以右肺下叶后基底段感染体积最大〔(31439.84±27135.99)mm3〕,感染占比最高〔占(28.23±14.82)%〕.相比人工胸部CT阅片,联影AI(uAI)辅助分析系统可自动检测病变并标记为可视化.结论 应用AI深度学习方法与人工阅片方式在分析胸部CT总体特征及异常征象方面无明显差异.但基于AI的胸部CT图像分析可直接以具体数据形式显示病灶分布特征和大小;另外uAI辅助分析系统可自动检测并标记病变位置,使病变部位更加直观.

Objective To initially discuss the characteristics of chest CT imaging of mild imported corona virus disease 2019(COVID-19)patients.Methods Using the artificial intelligence(AI)deep learning method,the imaging characteristics and clinical data of 52 patients with mild imported COVID-19 in some areas of Shenzhen were retrospectively analyzed,at the same time,two senior imaging doctors independently analyze and confirm each image result.The characteristics of chest CT with manual reading and AI were summarized,and the characteristics of CT analysis between manual reading and AI were compared.Results Compared with manual film reading,the AI chest CT analysis can automatically detect and mark the infected area,and at the same time quantify the infected area.The manual reading showed that all the patients had abnormal signs of lesions on chest CT,75.00%(39/52)of the lesions were scattered in two lungs,25.00%(13/52)of the lesions were single lobes or segments,67.31%(35/52)of the lesions were mostly involved in both lower lobes and subpleural;73.07%(38/52)of the CT showed ground glass density shadow,38.46%(20/52)of the lesions showed patchy exudation shadow,28.84%(15/52)of the lesions showed solid density shadow,which was consistent with the results of AI chest CT image analysis.The chest CT image analysis based on AI also showed that the most affected areas of pneumonia focus were double lower lobes,the infection volume[right lung(mm3):54857.13±23978.90 vs.46663.24±16898.62,left lung(mm3):34933.61±16400.66 vs.2348.57±1241.08]and infection ratio[right lung(%):7.19±5.40 vs.6.15±2.13,left lung(%):3.24±1.62 vs.0.19±0.11]were higher than upper lobes.In the affected lung segment,the infection volume[(31439.84±27135.99)mm3]and proportion[(28.23±14.82)%]of the right lower lobe posterior basal segment were the largest.Compared with manual reading,united AI(uAI)sysetm can automatically detect lesions and mark them as visualization.Conclusions There is no obvious difference in analyzing general characteristics and abnormal signs of chest CT using manual reading and AI deep learning.However,based on AI chest CT image analysis,it can directly display the distribution characteristics and size of the lesions in the form of specific data.In addition,the uAI intelligent auxiliary analysis system can automatically detect and mark the lesions,making the lesions more intuitive.