Evaluation of the Accuracy of Infer Read Software in Measuring the Volume of Pure Ground Glass Nodules in the Lung
-
摘要: 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。Abstract: 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.
-
Keywords:
- artificial intelligence /
- ground-glass nodule /
- volume /
- accuracy
-
-
![]()
图 1 pGGN的AI测量过程
（a）～（i）分别展示pGGN位于胸膜下、血管旁及肺实质3种情况时，AI自动识别并进行体积测量计算过程。（a）～（c）薄层CT扫描显示目标pGGN；（d）～（f）薄层CT扫描显示系统自动以方框标记目标pGGN；（g）～（i）薄层CT扫描显示结节与周围肺实质由系统自动勾勒边界，并自动计算结节体积（结节周围紧贴的血管、胸膜等结构由软件自动剔除）。
Figure 1. AI measurement process of pGGN
![]()
图 2 手动测量结节体积的过程
48岁女性，右肺中叶外侧段持续存在pGGN。（a）～（c）分别显示观察者使用PACS工具手动测量，精确提取pGGN过程；（d）显示由“推想”软件自动勾勒结节边缘并计算体积的成果。
Figure 2. Process of manually measuring nodule volume
![]()
图 3 AI与人工测量体积相关性散点图
Figure 3. Scatterplot of the correlation between AI and manual measurements
表 1 AI与人工测量值的差异性及相关性比较
Table 1 The differences and related comparisons between AI and manual measurements
性质 参数 AI测量 人工测量 P 差异性 P50（P25，P75） 264.59（158.08，544.77） 263.43（159.60，527.56） 0.703 相关性 相关系数r 1 0.981 0.000 注：AI与人工测量差异性比较采用Wilcoxon秩和检验，P=0.703>0.05，差异无统计学意义。二者相关性比较采用 Spearman分析，具有显著相关性；二者相关系数为0.981。 
下载: 导出CSV
表 2 AI与人工测量一致性比较
Table 2 The consistency comparisons between AI and manual measurements
项目 同类相关性 95％置信区间 使用真值0的F检验 下限 上限 值 自由度1 自由度2 显著性 单个测量 0.987 0.983 0.99 154.651 169 169 0 平均测量 0.994 0.991 0.995 154.651 169 169 0 注：AI与“金标准”一致性比较采用组内相关系数（ICC）一致性检验，结果显示：二者ICC值为0.987，一致性较好。
下载: 导出CSV
表 3 不同因素对AI测量误差的影响
Table 3 The effects of different factors on AI measurement errors
变量 分组 测量误差 Z P 体积 体积≤523.6 mm3（n=125） -0.553（-8.313，9.327） -0.769 0.442 体积>523.6 mm3（n=45） 3.847（-32.743，77.457） 位置 右肺上叶（n=49） 1.208（-4.947，14.561） 5.575 0.233 右肺中叶（n=21） -8.990（-22.598，12.897） 右肺下叶（n=27） 1.203（-8.542，16.477） 左肺上叶（n=49） -0.253（-12.017，12.932） 左肺下叶（n=24） -2.240（-20.935，12.388） 毗邻 肺实质（n=82） -0.745（-11.519，14.643） 0.438 0.803 贴近血管（n=29） 0.010（-6.762，15.200） 胸膜下（n=59） -0.253（-10.637，9.273）
下载: 导出CSV
-
[1] SUNG H, FERLAY J, SIEGEL R L, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries[J]. CA: A Cancer Journal for Clinicians, 2021, 71(3): 209−249. doi: 10.3322/caac.21660
[2] 吴林玉, 叶剑锋, 郑思思, 等. 应用人工智能对肺结节直径测量: 观察者内、观察者间差异[J]. 医学影像学杂志, 2020,30(1): 47−51. WU L Y, YE J F, ZHENG S S, et al. Differences in measurement of lung nodule diameter using artificial intelligence: Intraobserver and interobserver differences[J]. Journal of Medical Imaging, 2020, 30(1): 47−51. (in Chinese).
[3] MARTEN K, AUER F, SCHMIDT S, et al. Inadequacy of manual measurements compared to automated CT volumetry in assessment of treatment response of pulmonary metastases using RECIST criteria[J]. European Radiology, 2006, 16(4): 781−790. doi: 10.1007/s00330-005-0036-x
[4] PIANYKH O S, LANGS G, DEWEY M, et al. Continuous learning AI in radiology: Implementation principles and early applications[J]. Radiology, 2020, 297(1): 6−14. doi: 10.1148/radiol.2020200038
[5] REYES M, MEIER R, PEREIRA S, et al. On the Interpretability of artificial intelligence in radiology: Challenges and opposrtunities[J]. Radiology: Artificial Intelligence, 2020, 2(3): e190043−e190043. doi: 10.1148/ryai.2020190043
[6] HENSCHKE C I. Early lung cancer action project: Overall design and findings from baseline screening[J]. Cancer, 2000, 89(S11): 2474−2482. doi: 10.1002/1097-0142(20001201)89:11+<2474::AID-CNCR26>3.0.CO;2-2
[7] KANG S, KIM T H, SHIN J M, et al. Optimization of a chest computed tomography protocol for detecting pure ground glass opacity nodules: A feasibility study with a computer-assisted detection system and a lung cancer screening phantom[J]. PLoS One, 2020, 15(5): e0232688−e0232688. doi: 10.1371/journal.pone.0232688
[8] XU Y M, ZHANG T, XU H, et al. Deep learning in CT images: Automated pulmonary nodule detection for subsequent management using convolutional neural network[J]. Cancer Management and Research, 2020, 12: 2979−2992. doi: 10.2147/CMAR.S239927
[9] 朱江涛. 三维CT测量肺小结节体积的可重复性评价[D]. 苏州: 苏州大学, 2007: 20-24. [10] 孙炎冰, 陶广昱, 陈群慧, 等. 人工智能CT定量分析肺磨玻璃密度结节初探[J]. 中国医学计算机成像杂志, 2018,24(5): 383−387. doi: 10.3969/j.issn.1006-5741.2018.05.004 SUN Y B, TAO G Y, CHEN Q H, et al. Quantitative CT image features of ground glass opacity nodules: Initiative analysis with artificial intelligence[J]. Chinese Computed Medical Imaging, 2018, 24(5): 383−387. (in Chinese). doi: 10.3969/j.issn.1006-5741.2018.05.004
[11] 李梦琦, 韩融城, 宋文静,等. CT三维容积分析在实性肺结节恶性风险度评估中的价值[J]. 中国肺癌杂志, 2016,19(5): 279−285. doi: 10.3779/j.issn.1009-3419.2016.05.05 LI M Q, HAN R C, SONG W J, et al. Three dimensional volumetric analysis of solid pulmonary nodules on chest CT: Cancer risk assessment[J]. Chinese Journal of Lung Cancer, 2016, 19(5): 279−285. (in Chinese). doi: 10.3779/j.issn.1009-3419.2016.05.05
[12] 吴光耀, 伍建林. 基于CT的计算机辅助检测与诊断肺结节技术研究进展[J]. 中国医学影像技术, 2018,34(7): 1114−1117. WU G Y, WU J L. Research progresses of computer-aided detection and diagnosis based on CT in pulmonary nodules[J]. Chinese Journal of Medical Imaging Technology, 2018, 34(7): 1114−1117. (in Chinese).
[13] PETROU M, QUINT L E, NAN B, et al. Pulmonary nodule volumetric measurement variability as a function of CT slice thickness and nodule morphology[J]. American Journal of Roentgenology, 2007, 188(2): 306−312. doi: 10.2214/AJR.05.1063
[14] SON J Y, LEE H Y, KIM J H, et al. Quantitative CT analysis of pulmonary ground-glass opacity nodules for distinguishing invasive adenocarcinoma from non-invasive or minimally invasive adenocarcinoma: The added value of using iodine mapping[J]. European Radiology, 2016, 26(1): 43−54. doi: 10.1007/s00330-015-3816-y
[15] 侯永泉, 徐毛冶, 柴军, 等. 人工智能在肺小结节HRCT检出与诊断的研究进展[J]. 内蒙古医学杂志, 2021,53(2): 187−190. [16] 张怀瑢, 孙潇, 田兴仓, 等. 基于深度学习的人工智能诊断系统在不同重建参数下对肺结节的识别研究[J]. 临床放射学杂志, 2020,39(11): 2203−2206. ZHANG H R, SUN X, TIAN X C, et al. Deep Learning based artificial intelligence diagnosis system to identify pulmonary nodules under different reconstruction algorithms[J]. Journal of Clinical Radiology, 2020, 39(11): 2203−2206. (in Chinese).
[17] JAMES H T, XIANG L, QUANZHENG L, et al. Artificial intelligence and machine learning in radiology: Opportunities, challenges, pitfalls, and criteria for success[J]. Journal of the American College of Radiology, 2018, 15(3): 504−508. doi: 10.1016/j.jacr.2017.12.026
[18] ALI I, HART G R, GUNABUSHANAM G, et al. Lung nodule detection via deep reinforcement learning[J]. Frontiers in Oncology, 2018, 8(8): 108.
[19] HOSNY A, PARMAR C, QUACKENBUSH J, et al. Artificial intelligence in radiology[J]. Nature Reviews Cancer, 2018, 18(8): 500−510. doi: 10.1038/s41568-018-0016-5
[20] ARNAUD S A A, ALBERTO T, THOMAS D B, et al. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: The LUNA16 challenge[J]. Medical Image Analysis, 2017, 42: 1−13. doi: 10.1016/j.media.2017.06.015
[21] GOO J M, KIM K G, GIERADA D S, et al. Volumetric measurements of lung nodules with multi-detector row CT: Effect of changes in lung volume[J]. Korean Journal of Radiology, 2006, 7(4): 243−248. doi: 10.3348/kjr.2006.7.4.243
[22] ODA S, AWAI K, MURAO K, et al. Computer-aided volumetry of pulmonary nodules exhibiting ground-glass opacity at MDCT[J]. American Journal of Roentgenology, 2010, 194(2): 398−406. doi: 10.2214/AJR.09.2583
[23] HONDA O, SUMIKAWA H, JOHKOH T, et al. Computer-assisted lung nodule volumetry from multi-detector row CT: Influence of image reconstruction parameters[J]. European Journal of Radiology, 2006, 62(1): 106−113.
[24] 孙丹丹, 王亮, 许迪, 等. 不同重建算法对人工智能辅助肺结节检测效能的影响[J]. 生物医学工程与临床, 2021,25(2): 160−164. SUN D D, WANG L, XU D, et al. Effects of different reconstruction kernels on efficiency of artificial intelligence-assisted lung nodule detection[J]. Biomedical Engineering and Clinical Medicine, 2021, 25(2): 160−164. (in Chinese).
-
期刊类型引用(1)
1. 郭小皖，贾旭东，张丹青，贾德召，陈英敏，张淑倩，刘阳，时高峰. 肺部亚实性结节三维体积及体质量测量一致性分析. 温州医科大学学报. 2023(12): 974-979 . 
百度学术
其他类型引用(0)
计量
- 文章访问数: 442
- HTML全文浏览量: 131
- PDF下载量: 30
- 被引次数: 1
