-
目前,肺癌仍是全球癌症患者的主要病死原因之一,统计结果显示,2018年全球新增肺癌患者2 093 000例,占全球癌症患者总例数的12.22%[1]。其中,80%以上的肺癌都是非小细胞肺癌,早期非小细胞肺癌通常以孤立性肺结节(solitary pulmonary nodules,SPN)的形式发生,缺少临床症状[2]。
肺结核是由结核分支杆菌引起的一种常见的肉芽肿性炎症,在许多发展中国家由于卫生条件相对较差而得以流行,是继获得性免疫缺陷综合征之后由感染导致病死的第2大原因[3]。2019年,全球感染肺结核的患者约有1000万例,其中约130万例死于该病[4]。肺结核在18F-FDG PET/CT上亦可以表现为类似肺癌的高代谢SPN。
纹理分析是计算机辅助诊断的一个分支[5],其采用计算机图像分析技术对人眼无法识别的每个像素的空间分布和信号强度特点进行描述,通过特定的纹理参数量化评估相似病灶组织结构的异质性,从而对病灶进行鉴别诊断。本研究探讨MaZda纹理分析软件对18F-FDG PET/CT图像中的高代谢SPN鉴别诊断的增益价值。
-
由表1可知,肺癌与肺结核患者在性别及SPN长径间的差异均无统计学意义(均P>0.05);年龄的差异有统计学意义(P<0.05);肺癌组患者SPN的平均SUVmax高于肺结核组,且差异有统计学意义(P<0.05)。
组别 年龄(岁) 男/女(例) SPN长径
( cm)$\bar x \pm s, $ SUVmax
( )$\bar x \pm s $ 肺癌(n=63) 54(42~72) 38/25 2.09±0.64 9.51±4.65 肺结核(n=45) 47(35~64) 26/19 2.19±0.91 5.35±2.89 检验值 t=2.180 χ2=0.070 t=0.675 t=2.520 P值 0.034 0.791 0.511 0.014 注:SPN为孤立性肺结节;SUVmax为最大标准化摄取值 表 1 肺癌与肺结核患者一般临床资料的对比
Table 1. Comparison of basic information between lung cancer patients and tuberculosis patients
-
主观定性诊断高代谢SPN为肺癌或肺结核的错判率为26.9%(29/108)、灵敏度为93.7%(59/63)、特异度为35.9%(14/39)。以SUVmax作为鉴别诊断依据,ROC AUC为0.790(95%CI:0.674~0.906)(图1),当SUVmax临界值为5.3时,灵敏度为82.9%、特异度为64.0%、错判率为25.0%(27/108)。
-
SPN纹理分析结果显示,对于CT和PET图像,FPM纹理特征筛选联合非线性分类分析诊断的错判率均最低,分别为8.33%和1.85%(表2),比主观定性诊断和基于SUVmax的ROC曲线分析诊断的错判率均小,且差异均有统计学意义(χ2=10.800、27.457,均P<0.05)。与CT纹理分析结果相比,PET纹理分析鉴别肺癌与肺结核的错判率更低,且差异有统计学意义(χ2=4.694,P<0.05);然而,当主观定性诊断与基于SUVmax的ROC曲线分析结果比较时,错判率的差异无统计学意义(χ2=0.096,P>0.05)。图2为1例右肺上叶腺癌患者的CT、PET图像手动勾画SPN ROI的前后对比图。
纹理筛选方法 原始数据分析 主成分分析 线性分类分析 非线性分类分析 CT图像 Fisher系数 37.04(40/108) 37.04(40/108) 25.93(28/108) 11.11(12/108) POE+ACC 35.19(38/108) 35.19(38/108) 15.74(17/108) 9.26(10/108) MI 37.04(40/108) 37.04(40/108) 29.63(32/108) 10.19(11/108) FPM 40.74(44/108) 45.37(49/108) 16.67(18/108) 8.33(9/108) PET图像 Fisher系数 33.33(36/108) 24.07(26/108) 16.67(18/108) 9.26(10/108) POE+ACC 44.44(48/108) 44.44(48/108) 24.07(26/108) 7.41(8/108) MI 25.93(28/108) 22.22(24/108) 9.26(10/108) 3.70(4/108) FPM 44.44(48/108) 44.44(48/108) 7.41(8/108) 1.85(2/108) 注:CT为计算机体层摄影术;POE为分类错误概率;ACC为平均相关系数;MI为交互信息;FPM为Fisher系数+分类错误概率+平均相关系数+交互信息;PET为正电子发射断层显像术 表 2 各种纹理特征筛选方法联合分类分析方法诊断肺结核与肺癌高代谢孤立性肺结节错判率的对比(%)
Table 2. Comparison of the misclassified rates during various texture feature selection combined with feature classified analysis methods in identifying pulmonary tuberculosis and lung cancer hypermetabolic solitary pulmonary nodules (%)
图 2 右肺上叶腺癌患者(男性,54岁)的CT、PET图像手动勾画SPN感兴趣区的前后对比图
Figure 2. Comparison of the solitary pulmonary nodules' region of interest before and after manual delineation of CT and PET images in a patient (male, 54 years old) with adenocarcinoma of the right upper lobe
对FPM筛选出的PET图像的30个纹理特征参数分别行ROC曲线分析,由图3可见,灰度游程矩阵中的垂直方向长行程补偿、灰度游程矩阵中的135°方向长行程补偿和灰度共生矩阵中的逆差距纹理特征最具鉴别诊断价值,相应AUC分别为0.901(95%CI:0.820~0.983)、0.890(95%CI:0.803~0.977)和0.890(95%CI:0.804~0.976)。
18F-FDG PET/CT纹理分析在肺癌与肺结核的高代谢孤立性肺结节鉴别诊断中的增益价值
Complementary value of 18F-FDG PET/CT texture analysis in differential diagnosis of hypermetabolic lung cancer and tuberculosis nodules
-
摘要:
目的 探讨18F-氟脱氧葡萄糖(FDG) PET/CT纹理分析在肺癌与肺结核的高代谢孤立性肺结节(SPN)鉴别诊断中的增益价值。 方法 回顾性分析2017年1月至2020年6月在内蒙古赤峰市医院和北京大学肿瘤医院行18F-FDG PET/CT检查且结果表现为高代谢[最大标准化摄取值(SUVmax)≥2.5]的108例SPN患者的临床资料,其中男性68例、女性40例,年龄35~72岁,中位年龄50岁;肺结核患者45例(肺结核组)、肺癌患者63例(肺癌组)。所有患者均经组织病理学检查结果确诊。分析所有患者18F-FDG PET/CT图像SPN的良恶性(主观定性诊断),并计算错判率、灵敏度和特异度。采用MaZda纹理分析软件分别对CT和PET图像中的SPN横断面最大层面及相邻上下两层图像手动勾画ROI并提取纹理特征参数。分别采用Fisher系数、分类错误概率+平均相关系数、交互信息以及三者联合的方法(FPM)筛选具有鉴别意义的纹理特征。对筛选出的纹理特征进行原始数据、主成分、线性分类和非线性分类的分析,对SPN的良恶性进行鉴别,以错判率评价其鉴别效能。计量资料的组间比较采用独立样本t检验;计数资料的组间比较采用χ2检验。对错判率最低的各纹理特征参数分别进行受试者工作特征(ROC)曲线分析,筛选最具鉴别意义的前3位纹理特征。 结果 肺癌组与肺结核组患者在年龄[54(42~72)岁对47(35~64)岁]、SUVmax[(9.51±4.65)对(5.35±2.89)]间的差异均有统计学意义(t=2.180、2.520,均P<0.05);在性别、SPN长径间的差异均无统计学意义(χ2=0.070,t=0.675,均P>0.05)。主观定性诊断高代谢SPN良恶性的错判率为26.9%(29/108)、灵敏度为93.7%(59/63)、特异度为35.9%(14/39)。基于SUVmax的ROC曲线分析,SUVmax临界值为5.3时错判率为25.0%(27/108),其与主观定性诊断的错判率差异无统计学意义(χ2=0.096,P>0.05)。CT和PET图像基于FPM联合非线性分类分析诊断的错判率最低,分别为8.33%和1.85%,两者间的差异有统计学意义(χ2=4.694,P<0.05);其与主观定性诊断和基于SUVmax的ROC曲线分析比较,错判率的差异均有统计学意义(χ2=10.800、27.457,均P<0.05)。最具鉴别意义的前3位纹理特征分别为灰度游程矩阵中的垂直方向长行程补偿、灰度游程矩阵中的135°方向长行程补偿和灰度共生矩阵中的逆差距。 结论 MaZda纹理分析鉴别高代谢SPN良恶性的诊断效能较主观诊断高,在肺癌与肺结核的高代谢SPN的鉴别诊断中具有增益价值。 -
关键词:
- 孤立性肺结节 /
- 氟脱氧葡萄糖F18 /
- 正电子发射断层显像术 /
- 体层摄影术,X线计算机 /
- 肺肿瘤 /
- 结核,肺 /
- 纹理分析
Abstract:Objective To explore the complementary value of 18F-fluorodeoxyglucose (FDG) PET/CT texture analysis in the differential diagnosis of solitary lung cancer and tuberculosis nodules with hypermetabolic solitary pulmonary nodules (SPN). Methods A total of 108 patients with hypermetabolic SPN (maximum standard uptake value (SUVmax)≥2.5) who were examined by 18F-FDG PET/CT were recruited retrospectively in Chifeng Municipal Hospital of Inner Mongolia and Beijing Cancer Hospital, Beijing Institute for Cancer Research from January 2017 to June 2020. The patients consisted of 68 males and 40 females aged 35 to 72 years, with a median age of 50 years. Forty-five patients had tuberculosis (tuberculosis group), and 63 patients had lung cancer (lung cancer group). All the nodules were confirmed by pathological examination. The benign and malignant SPN of the 18F-FDG PET/CT images of all patients were analyzed (subjective qualitative diagnosis), and misjudgment rate, sensitivity, and specificity were calculated. MaZda texture analysis software was used to delineate the region of interest manually on the maximum cross section and the adjacent upper and lower layers of the lung nodule on the CT and PET images to extract texture feature parameters. Fisher coefficient, classification error probability + average correlation coefficients, mutual information, and their combination (FPM) were used to filter texture features with differential diagnostic significance. The software program was also used for raw data analysis, principal component analysis, linear discriminant analysis, and nonlinear discriminant analysis to discriminate benign or malignant nodules. The discriminating efficacy was evaluated by misclassified rate. Independent sample t test was used to compare the measurement data between groups. Chi-square test was used to compare the count data between groups. Receiver operating characteristic (ROC) curve analysis was carried out for each texture feature parameter of the lowest misclassified rate to screen the top three most discriminative texture features. Results The differences in age (54(42–72) years old vs. 47(35–64) years old) and SUVmax ((9.51±4.65) vs. (5.35±2.89)) were statistically significant between the tuberculosis group and lung cancer group (t=2.180, 2.520; both P<0.05). No statistically significant differences in gender and long diameter (χ2=0.070, t=0.675; both P>0.05) were found. The misclassified rate of the subjective qualitative diagnosis of high metabolic SPN was 26.9% (29/108). The sensitivity was 93.7% (59/63), and the specificity was 35.9% (14/39). ROC curve analysis was conducted based on SUVmax. When SUVmax had the best cut-off value of 5.3, the misclassified rate was 25.0% (27/108). No significant difference was found when compared with subjective qualitative diagnosis (χ2=0.096, P>0.05). The CT and PET images based on the combination of FPM feature selection and nonlinear discriminant analysis had lower misclassification rates of 8.33% and 1.85%, respectively, and the difference between them was statistically significant (χ2=4.694, P<0.05). Compared with subjective qualitative diagnosis and ROC curve analysis based on SUVmax, the differences were statistically significant (χ2=10.800, 27.457; both P<0.05). The most discriminative texture features were vertical long-run emphasis and 135° long-run emphasis in the gray-scale run length matrix and inverse difference moment in the gray-level co-occurrence matrix. Conclusion MaZda texture analysis has higher diagnostic efficiency than subjective judgment in identifying benign and malignant hypermetabolic SPN nodules and thus complementary value in the differential diagnosis of hypermetabolic lung cancer and tuberculosis nodules. -
表 1 肺癌与肺结核患者一般临床资料的对比
Table 1. Comparison of basic information between lung cancer patients and tuberculosis patients
组别 年龄(岁) 男/女(例) SPN长径
( cm)$\bar x \pm s, $ SUVmax
( )$\bar x \pm s $ 肺癌(n=63) 54(42~72) 38/25 2.09±0.64 9.51±4.65 肺结核(n=45) 47(35~64) 26/19 2.19±0.91 5.35±2.89 检验值 t=2.180 χ2=0.070 t=0.675 t=2.520 P值 0.034 0.791 0.511 0.014 注:SPN为孤立性肺结节;SUVmax为最大标准化摄取值 表 2 各种纹理特征筛选方法联合分类分析方法诊断肺结核与肺癌高代谢孤立性肺结节错判率的对比(%)
Table 2. Comparison of the misclassified rates during various texture feature selection combined with feature classified analysis methods in identifying pulmonary tuberculosis and lung cancer hypermetabolic solitary pulmonary nodules (%)
纹理筛选方法 原始数据分析 主成分分析 线性分类分析 非线性分类分析 CT图像 Fisher系数 37.04(40/108) 37.04(40/108) 25.93(28/108) 11.11(12/108) POE+ACC 35.19(38/108) 35.19(38/108) 15.74(17/108) 9.26(10/108) MI 37.04(40/108) 37.04(40/108) 29.63(32/108) 10.19(11/108) FPM 40.74(44/108) 45.37(49/108) 16.67(18/108) 8.33(9/108) PET图像 Fisher系数 33.33(36/108) 24.07(26/108) 16.67(18/108) 9.26(10/108) POE+ACC 44.44(48/108) 44.44(48/108) 24.07(26/108) 7.41(8/108) MI 25.93(28/108) 22.22(24/108) 9.26(10/108) 3.70(4/108) FPM 44.44(48/108) 44.44(48/108) 7.41(8/108) 1.85(2/108) 注:CT为计算机体层摄影术;POE为分类错误概率;ACC为平均相关系数;MI为交互信息;FPM为Fisher系数+分类错误概率+平均相关系数+交互信息;PET为正电子发射断层显像术 -
[1] Bray F, Ferlay J, Soerjomataram I, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries[J]. CA Cancer J Clin, 2018, 68(6): 394−424. DOI: 10.3322/caac.21492. [2] Rzechonek A, Grzegrzolka J, Blasiak P, et al. Correlation of expression of tenascin C and blood vessel density in non-small cell lung cancers[J]. Anticancer Res, 2018, 38(4): 1987−1991. DOI: 10.21873/anticanres.12436. [3] Kim I, Lee JS, Kim SJ, et al. Double-phase 18F-FDG PET-CT for determination of pulmonary tuberculoma activity[J]. Eur J Nucl Med Mol Imaging, 2008, 35(4): 808−814. DOI: 10.1007/s00259-007-0585-0. [4] World Health Organization. Global tuberculosis report 2020[EB/OL]. [2020-10-31]. https://apps.who.int/iris/handle/10665/336069. [5] Holli-Helenius K, Salminen A, Rinta-Kiikka I, et al. MRI texture analysis in differentiating luminal A and luminal B breast cancer molecular subtypes—a feasibility study[J/OL]. BMC Med Imaging, 2017, 17(1): 69[2020-10-31]. https://bmcmedimaging.biomedcentral.com/articles/10.1186/s12880-017-0239-z. DOI: 10.1186/s12880-017-0239-z. [6] Jeong YJ, Yi CA, Lee KS. Solitary pulmonary nodules: detection, characterization, and guidance for further diagnostic workup and treatment[J]. AJR Am J Roentgenol, 2007, 188(1): 57−68. DOI: 10.2214/AJR.05.2131. [7] Lin JZ, Zhang L, Zhang CY, et al. Application of gemstone spectral computed tomography imaging in the characterization of solitary pulmonary nodules: preliminary result[J]. J Comput Assist Tomogr, 2016, 40(6): 907−911. DOI: 10.1097/RCT.0000000000000469. [8] Strzelecki M, Szczypinski P, Materka A, et al. A software tool for automatic classification and segmentation of 2D/3D medical images[J]. Nucl Instrum Methods Phys Res Sect A: Accelerat, Spectrom, Detect Assoc Equip, 2013, 702: 137−140. DOI: 10.1016/j.nima.2012.09.006. [9] Teramoto A, Tsukamoto T, Kiriyama Y, et al. Automated classification of lung cancer types from cytological images using deep convolutional neural networks[J]. Biomed Res Int, 2017, 2017: 4067832. DOI: 10.1155/2017/4067832. [10] Xiao N, Qiang Y, Bilal Zia M, et al. Ensemble classification for predicting the malignancy level of pulmonary nodules on chest computed tomography images[J]. Oncol Lett, 2020, 20(1): 401−408. DOI: 10.3892/ol.2020.11576. [11] Niyonkuru A, Chen XM, Bakari KH, et al. Evaluation of the diagnostic efficacy of 18F-fluorine-2-deoxy-D-glucose PET/CT for lung cancer and pulmonary tuberculosis in a tuberculosis-endemic country[J/OL]. Cancer Med, 2020, 9(3): 931−942[2020-10-31]. https://onlinelibrary.wiley.com/doi/10.1002/cam4.2770. DOI: 10.1002/cam4.2770. [12] Thawani R, McLane M, Beig N, et al. Radiomics and radiogenomics in lung cancer: a review for the clinician[J]. Lung Cancer, 2018, 115: 34−41. DOI: 10.1016/j.lungcan.2017.10.015. [13] Lee G, Lee HY, Park H, et al. Radiomics and its emerging role in lung cancer research, imaging biomarkers and clinical management: state of the art[J]. Eur J Radiol, 2017, 86: 297−307. DOI: 10.1016/j.ejrad.2016.09.005. [14] Labby ZE, Armato Ⅲ SG, Dignam JJ, et al. Lung volume measurements as a surrogate marker for patient response in malignant pleural mesothelioma[J]. J Thorac Oncol, 2013, 8(4): 478−486. DOI: 10.1097/JTO.0b013e31828354c8. [15] Lee JH, Lee HY, Ahn MJ, et al. Volume-based growth tumor kinetics as a prognostic biomarker for patients with EGFR mutant lung adenocarcinoma undergoing EGFR tyrosine kinase inhibitor therapy: a case control study[J]. Cancer Imaging, 2016, 16: 5. DOI: 10.1186/s40644-016-0063-7. [16] Fogel I, Sagi D. Gabor filters as texture discriminator[J]. Biol Cybern, 1989, 61(2): 103−113. DOI: 10.1007/BF00204594. [17] Bhargava R, Madabhushi A. Emerging themes in image informatics and molecular analysis for digital pathology[J]. Annu Rev Biomed Eng, 2016, 18: 387−412. DOI: 10.1146/annurev-bioeng-112415-114722. [18] Szczypiński PM, Strzelecki M, Materka A, et al. MaZda—a software package for image texture analysis[J]. Comput Methods Programs Biomed, 2009, 94(1): 66−76. DOI: 10.1016/j.cmpb.2008.08.005. [19] Mannil M, Burgstaller JM, Held U, et al. Correlation of texture analysis of paraspinal musculature on MRI with different clinical endpoints: lumbar Stenosis Outcome Study (LSOS)[J]. Eur Radiol, 2019, 29(1): 22−30. DOI: 10.1007/s00330-018-5552-6. [20] Hodgdon T, Thornhill RE, James ND, et al. CT texture analysis of acetabular subchondral bone can discriminate between normal and cam-positive hips[J]. Eur Radiol, 2020, 30(8): 4695−4704. DOI: 10.1007/s00330-020-06781-1. [21] Holli-Helenius K, Salminen A, Rinta-Kiikka I, et al. MRI texture analysis in differentiating luminal A and luminal B breast cancer molecular subtypes—a feasibility study[J/OL]. BMC Med Imaging, 2017, 17(1): 69[2020-10-31]. https://bmcmedimaging.biomedcentral.com/articles/10.1186/s12880-017-0239-z. DOI: 10. 1186/s12880-017-0239-z. [22] Hui TCH, Chuah TK, Low HM, et al. Predicting early recurrence of hepatocellular carcinoma with texture analysis of preoperative MRI: a radiomics study[J]. Clin Radiol, 2018, 73(12): 1056.e11−1056.e16. DOI: 10.1016/j.crad.2018.07.109. [23] 鄂林宁, 张娜, 王荣华, 等. 计算机体层摄影术纹理分析对孤立性肺结节良恶性鉴别诊断的价值[J]. 中华肿瘤杂志, 2018, 40(11): 847−850. DOI: 10.3760/cma.j.issn.0253-3766.2018.11.010.
E LN, Zhang N, Wang RH, et al. Comparative analysis of computed tomography texture features between pulmonary inflammatory nodules and lung cancer[J]. Chin J Oncol, 2018, 40(11): 847−850. DOI: 10.3760/cma.j.issn.0253-3766.2018.11.010.