InferRead软件测量肺纯磨玻璃结节体积的准确性评价

Evaluation of the Accuracy of Infer Read Software in Measuring the Volume of Pure Ground Glass Nodules in the Lung

目的:探讨人工智能(AI)肺结节检测软件对肺纯磨玻璃结节(pGGN)体积自动测量的准确性及测量误差的影响因素。方法:选择2021年1月1日至31日在本院行常规胸部CT扫描的患者90例,共计170个pGGN病灶。将CT扫描原始数据(含1 mm薄层图像)传送至AI服务器进行肺结节体积自动测量并记录其测量数据;由两名资深胸部影像诊断医师以手动进行pGGN逐层测量相加得出体积数值,并以3次测量平均值作为“金标准”数据与AI测量结果进行比较;并分析pGGN体积大小、位置、毗邻等因素对AI测量误差的影响。采用SPSS 26.0进行统计分析。结果:本研究90例患者中共计170个pGGN,右肺上叶者49个(28.82%),右肺中叶者21个(12.35%),右肺下叶者27个(15.89%),左肺上叶者49个(28.82%),左肺下叶者24个(14.12%)。在pGGN的毗邻关系中,完全位于肺实质内无毗邻关系者82个(48.24%),贴近血管者29个(17.06%),贴近胸膜者59个(34.70%)。①两名观察者之间、同一观察者不同时间点对pGGN手动测量的体积数值均无差异;②对相同pGGN体积的测量,使用AI自动测量与人工手动测量的结果亦无差异,且二者相关性很高(r=0.981),一致性也很高(ICC值为0.987);③pGGN的体积大小、发生位置、毗邻关系对AI体积测量的误差均无统计学意义。结论:InferRead肺结节检测软件对肺pGGN三维体积测量具有良好的准确性,可适用于临床肺结节诊断与相关研究。

Objective: To discuss the accuracy of the automatic measurement of the volume of pure ground glass nodules(pGGN) by the artificial intelligence pulmonary nodule detection software and the influencing factors of the measurement error. Methods: 170 pGGNs from 90 patients who underwent routine chest CT scan from January 1 to 31, 2021 in our hospital were selected in this retrospective study. The original CT scan data(including 1 mm thin-slice images) was sent to the AI server of Inference Technology to automatically measure the volume of lung nodules and record the measurement data. Two senior chest imaging diagnosticians manually carried out pGGNs layer-by-layer measurement and added the volume value, then took the average of three measurements as the "gold standard" data to compare with the AI measurement results, and then analyzed the influence of pGGN location,size, proximity and other factors on the AI measurement error. SPSS 26.0 was used for statistical analysis. Results:Among the total 170 pGGNs in the 90 patients in this study, 49(28.82%) were in the right upper lobe, 21(12.35%)were in the right middle lobe, 27(15.88%) were in the right lower lobe, and left upper lobe 49 patients(28.82%)were involved, and 24 patients were in the lower lobe of the left lung(14.12%). Among the adjacent relationships of pGGN, 82(48.24%) were completely located in the lung parenchyma without adjacency, 29(17.06%) were close to the blood vessel, and 59(34.70%) were close to the pleura.(1) There was no statistically significant difference in the volume values of pGGNs between two observers and the same observer at different time points.(2) For the measurement of the same pGGN, there was no statistically significant difference between the results of automatic measurement by AI and manual measurement and the correlation between the two was quite high(r = 0.981), and the agreement was also very high(ICC value is 0.987).(3) The size, location, and adjacent relationship of pGGN lesions were not statistically significant for the error of AI volume measurement.Conclusions: The InferRead lung nodule measurement software shows high accuracy in the measurement of lung pGGN three-dimensional volume, which can be applied in clinical lung nodule diagnosis and related research.(2) For the measurement of the same pGGN volume, there was no difference between the results of automatic measurement by AI and manual measurement, and the correlation between them was quite high(r = 0.981), and the consistency was high(ICC value is 0.987).(3) The volume size, occurrence position and adjacent relationship of pGGN have no statistical significance on the error of AI volume measurement. Conclusion: The InferRead lung nodule detection software shows high accuracy in measuring the three-dimensional volume of lung pGGN, and can be applied to clinical diagnosis and related research of lung nodules.