人工智能辅助诊断系统在新型冠状病毒肺炎诊断中的初步应用

Preliminary application of AI diagnosis system in the diagnosis of the novel coronavirus infected pneumonia

目的:评估人工智能(AI)辅助诊断系统对新型冠状病毒肺炎(NCIP)的诊断价值。方法:回顾性分析26例NCIP患者的临床资料与CT图像。利用AI系统新冠肺炎诊断模块自动识别肺炎病变的数量、测量病变体积、计算病变体积所占肺叶的百分比,得出病变的疑似新冠肺炎的概率。评估该AI系统识别病变数量的准确性和病变范围的准确性。结果:26例NCIP患者AI自动检测出215处病变,其中76.9%(20/26)的患者AI识别的病变数量高于医生评估的数量。23.1%(6/26)的患者病变体积小于10 cm^3,38.5%(10/26)的患者病变体积为10~100 cm^3,38.5%(10/26)的患者病变体积大于100 cm^3。57.7%(15/26)的患者病变体积百分比小于10%,23.1%(6/26)的患者体积百分比为10%~25%,15.4%(4/26)的患者体积百分比为25%~50%,3.8%(1/26)的患者体积百分比大于50%。34.6%(9/26)的患者NCIP的疑似概率大于50%,其中仅11.5%(3/26)的患者疑似概率为99%,65.4%(17/26)的患者疑似概率小于25%。61.5%(16/26)的患者肺部一些较小的病变未被AI识别,38.5%(10/26)的患者两个相邻病变识别为同一病变,42.3%(11/26)的患者肺部伪影被识别为病变,61.5%(16/26)的患者肺部正常结构被识别为病变,30.8%(8/26)的患者肺内的一些其他病灶被识别为NCIP病变,46.2%(12/26)的患者小病灶被识别为肺结节。88.5%(23/26)的患者部分病变边缘正常的肺组织被识别为病变;61.5%(16/26)的患者部分病变边缘未被识别为病变。结论:AI辅助诊断系统能迅速识别CT影像中的肺炎病变,并进行体积测量,给出NCIP的疑似概率,帮助医生快速识别高危人群,判断病情的严重程度以及疗效,节省时间和精力,达到精准防控的效果。

Objective To evaluate the diagnostic value of artificial intelligence(AI)assisted diagnosis system for new coronavirus pneumonia(NCIP).Methods The clinical data and CT images of 26 NCIP patients were retrospectively analyzed.The new coronary pneumonia diagnosis module of the AI system is used to automatically identify the number of pneumonia lesions,measure the volume of the lesion,and calculate the percentage of the lesion volume in the lung lobe to obtain the probability of suspected new coronary pneumonia.Evaluate the accuracy of the AI system in identifying the number of lesions and the accuracy of the scope of the lesion.Results The AI system automatically detected 215 lesions of 26 NCIP patients.In 76.9%(20/26)of the patients AI has recognized more lesions than the doctor did.23.1%(6/26)of the patients had a lesion volume less than 10 cm^3,38.5%(10/26)had a lesion volume of 10-100 cm^3,and 38.5%(10/26)had a lesion volume greater than 100 cm^3.57.7%(15/26)patients have lesion volume percentage less than 10%,23.1%(6/26)patients have a volume percentage of 10%~25%,15.4%(4/26)patients have a volume percentage of 25%~50%,and only 3.8%(1/26)of the patient have a volume percentage greater than 50%.34.6%(9/26)patients have a suspected probability of NCIP greater than 50%,of which only 11.5%(3/26)patients have a suspected probability of 99%,and 65.4%(17/26)patients have a suspected probability of less than 25%.61.5%(16/26)of the patients have some smaller lesions unrecognized by AI,38.5%(10/26)of the patients have two adjacent lesions that was mistakenly recognized as the same lesion,and there are also 42.3%(11/26)of the patients whose lungs partial artifacts were identified as lesions,61.5%(16/26)of the patients whose normal lung structures were identified as lesions,30.8%(8/26)of the patients’whose lungs were identified as NCIP lesions,and 46.2%(12/26)of the patients whose small lesions were identified as lung nodules.In 88.5%(23/26)patients,some lung tissues with normal edges of the lesion were recognized as lesions;in 61.5%(16/26)of the patients the lesion edge was not recognized.Conclusion The AI-assisted diagnosis system can quickly identify pneumonia lesions in CT images,and measure the volume,give the suspected probability of NCIP,help doctors quickly identify high-risk groups,determine the severity and efficacy of the disease,save time and energy,and achieve precise prevention and control Effect.