人工智能对肺结核患者病变检出及定性诊断价值研究

Clinical value of Artificial Intelligence in detecting and analyzing lesions in patients with pulmonary tuberculosis

目的对比肺结节检测分析系统和放射科医师对肺结核患者病变的检出情况,探讨人工智能(Artificial Intelligence,AI)对肺结核患者病变检出的价值。方法收集北京大学深圳医院2016年12月至2018年3月行胸部高分辨CT平扫检查的32例肺结核病例并分为A、B两组,分别评价两组的CT图像。A组采用σ-Discover/Lung肺结节检测分析系统0.2版本,分析获得病变检测结果;B组为2名胸部影像医师分析并获得病变检测结果。统计分析A、B两组对肺结核患者病变的检出情况,回顾性分析CT图像,经2名经验丰富的胸部影像科医师阅片,剔除考虑干酪性肺炎等大片状阴影的病例,采用σ-Discover/Lung肺结节检测分析系统分析病变良恶性概率,以大于50%作为高恶性判断标准。结果本研究共纳入32例肺结核患者,共413处病灶。A组一共检出380处病灶,其中误检10处,漏检43处;B组一共检出407处病灶,其中无误检,漏检6处病变。A组对肺结核患者病变的检出率为89.83%,误检率为2.63%,漏诊率为10.17%;B组对病变的检出率为98.55%,误检率为0%,漏诊率为1.42%。肺结节检测分析系统判断一共217处病变恶性概率大于50%,57.1%的肺良性病变分析为高恶性概率。结论肺结节检测分析系统对肺结核患者病变的检出具有较好的能力,但检出能力不如放射科医师,漏诊率较高,且目前对肺结核患者病变的良恶性分析准确率较低。

Objective To compare the detection abilities of pulmonary nodule detection and analysis system and radiologists in the diagnosis of pulmonary tuberculosis lesions. To explore the contribution of artificial intelligence to the detection of lesions in tuberculosis patients. Methods A retrospective study was conducted to analyze the computed tomography images of tuberculosis patients from December 2016 to March 2018 in Peking University Shenzhen hospital. Firstly,CT images were retrospectively analyzed by two methods and divided into two groups,A and B. Group A was analyzed by σ-Discover/Lung pulmonary nodule detection and analysis system 1.0. Group B was analyzed by two chest radiologists. The abilities of lesions detection in tuberculosis patients in groups A and B were compared. Secondly,after reading by two experienced chest radiologists,patients with malignant pulmonary nodules or masses were excluded,and σ-Discover/Lung pulmonary nodule detection and analysis system was used to analyze the probabilities of malignant lesions(more than 50% as a high malignant criterion). Results A total of 32 patients with pulmonary tuberculosis were enrolled in the study,with a total of 413 lesions. A total of 380 lesions were detected in group A,including 10 misdetections and 43 missed examinations. A total of 407 lesions were detected in group B and with no false detections and 6 missing lesions. The detection rate of lesions in group A was 89.83%,the rate of false detection was 2.63%,and the rate of missing diagnosis was 10.17%. The detection rate of lesions in group B was 98.55%,the rate of false detection was 0%,and the rate of missing diagnosis was 1.42%. The pulmonary nodule detection and analysis system judged that the total malignant probability of 217 lesions was greater than 50%,and 57.1% of benign lung lesions were analyzed for high malignant probability. Conclusions The pulmonary nodule detection and analysis system has a good ability to detect the lesions of pulmonary tuberculosis patients,but the detection ability is not as good as radiologists. The rate of missing diagnosis is high,and the current accuracy of diagnosis of benign lesions of tuberculosis patients is low.