基于3D增强CT影像组学的肾癌亚型三分类预测模型

张海捷; 殷夫; 陈梦林; 漆安琪; 杨丽洋; 崔维维; 杨姗姗; 文戈

doi:10.12122/j.issn.1674-4500.2021.03.03

基于3D增强CT影像组学的肾癌亚型三分类预测模型

doi: 10.12122/j.issn.1674-4500.2021.03.03

1.
深圳大学第一附属医院 PET/CT中心，广东深圳 518052
2.
深圳大学信息工程学院，广东深圳 518052
3.
南方医科大学南方医院影像系，广东广州 510515

基金项目:

广东省自然科学基金 2020A151501046

详细信息

作者简介:
张海捷，硕士，主治医师，E-mail: zhjszey@163.com

通讯作者:
文戈，博士，主任医师，E-mail: m13360022166@163.com

计量
- 文章访问数: 791
- HTML全文浏览量: 345
- PDF下载量: 65
- 被引次数: 0
出版历程
- 收稿日期: 2021-04-03
- 刊出日期: 2021-05-20

A three categories prediction model for renal cell carcinoma subtype based on 3D enhanced CT radiomics

1.
PET/CT Center, First Affiliated Hospital of Shenzhen University, Shenzhen 518052, China
2.
School of Information Engineering, Shenzhen University, Shenzhen 518052, China
3.
Department of Imaging, Nanfang Hospital, Southern Medical University, Guangzhou 510515, China

摘要

摘要: 目的探讨可靠的基于3D多期增强CT影像组学特征的肾癌亚型三分类预测模型。方法210例肾细胞癌患者（透明细胞癌143例，乳头状癌25例，嫌色细胞癌29例，其他亚型的肾细胞癌13例）被纳入研究。使用ITK-SNAP软件，获取患者的3D增强CT病灶分割图像，使用PyRadiomics计算平台进行特征提取，使用集成学习分层bagging法来筛选特征和构建肾细胞癌亚型三分类预测模型：首先用100次5折交叉验证将模型分为训练集和测试集，然后将Lasso回归作为基学习器对影像组学特征的进行筛选，最后使用logistic回归作为基学习器进行建模和校正。根据不同期像的CT图像，构建平扫期模型、皮髓质期模型、实质期模型、排泄期模型和全期模型。使用准确率、精确率、敏感度和Kappa值评估测试集上不同期像预测模型的性能。结果每期CT图像中提取到了105个影像组学特征。在5个模型中，全期模型的预测效能最好，准确率为0.81，AUC为0.85；精确度为0.717；敏感度为0.799，kappa值为0.679。全期模型的影像组学特征中，有4个皮髓质期特征、3个实质期特征、1个排泄期特征和1个平扫期特征，且与其他4个单期模型中的特征没有重叠。在4个单期模型中，实质期模型的性能最好，准确性0.786，精确度0.689，敏感度0.734，AUC 0.811，Kappa值0.532；皮髓质期模型和排泄期模型的性能相似，但是排泄期模型的Kappa值0.285，明显低于皮髓质期的Kappa值0.446。平扫期模型的性能最差，AUC为0.693。结论基于3D多期增强CT影像组学特征的全期模型是区分肾细胞癌亚型的可靠和有效的方法。
- 3D成像 /
- 肾细胞癌亚型 /
- 影像组学 /
- 集成学习 /
- 机器学习 /
- 三分类
Abstract: ObjectiveTo construct an effective and reliable three categories prediction model to distinguish RCC subtypes based on 3D multi-phase enhanced computed tomography (CT) radiomic features (RFs).MethodsA total of 210 RCC were included in this study (143 clear cell, 25 papillary, 29 chromophobe, and 13 other RCC). The 3D multi-phase enhanced CT-based RFs were used to construct a prediction model. CT included a non-contrast phase (NCP), cortico-medullary phase (CMP), parenchyma phase (PP), excretory phase (EP), and all-phase (ALL-P), which contains all the single-phase information. The ensemble learning stratified bagging method was used to predict the RCC subtype by using LASSO regression and a 1 vs. rest logistic regression algorithm. Five-fold and external stratified cross-validation was used to assess the performance of the different prediction models.ResultsThere were 105 RFs extracted from each single phase of the CT scan. The 4 CMP, 3 PP, 1 EP, and 1 NCP RF were selected in the ALL-P model, and these RFs was no overlap with the other 4 single-phase models. The prediction efficiency of ALL-P was the best, with a diagnostic accuracy of 0.81 [area under the receiver operating characteristic curve (AUC)=0.853, precision=0.717, sensitivity=0.799, kappa=0.679). Among the four single phase models, the PP model had the best performance, with accuracy of 78.3% (AUC=0.811, precision=0.689, sensitivity=0.735, kappa=0.532). The performance of the CMP model was similar to that of the EP model, but the kappa value of the EP model was 0.285, which was significantly lower than that of the CMP model (0.446). The performance of the model in NCP was the worst, AUC was 0.693.ConclusionThe ALL-P prediction model based on 3D CT RFs is an effective and reliable method for distinguishing RCC subtypes.
- 3D imaging /
- renal cell carcinoma subtype /
- radiomic features /
- ensemble learning /
- machine learning /
- three categories

HTML全文

图 1 患者入组流程图

Figure 1. Flow chart of patient enrollment.

下载: 全尺寸图片幻灯片

图 2 皮髓质期分割图像示例

Figure 2. CT image segmentation in a cortico-medullary phase enhanced CT scan.

下载: 全尺寸图片幻灯片

图 3 集成学习特征筛选流程图

Figure 3. Flow chart of the ensemble learning bagging method of feature selection.

下载: 全尺寸图片幻灯片

图 4 集成学习模型构建流程图

Figure 4. Flow chart of the ensemble learning bagging method of model building.

下载: 全尺寸图片幻灯片

表 1 不同肾癌亚型患者的一般情况

Table 1. General situation of patients with different subtypes of renal cell carcinoma [n(%)]

因素	肾细胞癌				P
因素	透明细胞癌(n=143)	乳头状肾癌(n=25)	嫌色细胞癌(n=29)	其他类型肾癌(n=13)	P
性别					0.881
男	92(64.3)	18(72.0)	13(44.8)	7(53.8)
女	51(35.7)	7(28.0)	16(55.2)	6(46.2)
年龄（岁，Mean±SD）	53.2±11.58	51.3±12.59	54.14±13.76	52.1±19.88	0.327
分布					0.772
右侧	62(43.4)	15(60.0)	16(55.2)	5(38.5)
左侧	81(56.6)	10(40.0)	13(44.8)	8(61.5)

下载: 导出CSV

表 2 各期模型的影像组学特征和LASSO系数

Table 2. Radiomic features and lasso coefficient of each phase model

模型	影像组学特征名称	特征类別	LASSO系数
NCP	Volume	Shape	0.098130
	Major Axis	Shape	0.104464
	Least Axis	Shape	0.12697
	Maximum 2D Diameter Column	Shape	0.054351
	Large Dependence Emphasis	Gldm	0.123407
	Large Dependence Low Gray Level Emphasis	Gldm	0.098824
	Large Dependence High Gray Level Emphasis	Gldm	0.123407
	Kurtosis	First order	0.085602
	Run Length Non Uniformity	glrlm	0.080397
	Long Run High Gray Level Emphasis	glrlm	0.123407
CMP	Large Dependence Emphasis	Gldm	0.172186
	Cluster Prominence	Glcm	0.172011
	Total Energy	First order	0.101716
	Gray Level Non Uniformity	Glrlm	0.149354
	Size Zone Non Uniformity	Glszm	0.153395
	Large Area Low Gray Level Emphasis	Glszm	0.133549
	Large Area High Gray Level Emphasis	Glszm	0.132384
PP	Dependence Non Uniformity	Gldm_	0.088395
	Gray Level Non Uniformity	Gldm_	0.168465
	Large Dependence Emphasis	Gldm_	0.082350
	Total Energy	First order	0.149260
	Long Run High Gray Level Emphasis	Glrlm	0.088069
	Zone Variance	Glszm	0.147044
	Gray Level Non Uniformity	Glszm	0.144033
	Busyness	Ngtdm	0.132384
EP	Volume	Shape	0.092024
	Gray Level Non Uniformity	Gldm	0.099262
	Large Dependence Emphasis	Gldm	0.128963
	Total Energy	First order	0.082465
	Gray Level Non Uniformity	Glrlm	0.105180
	Run Length Non Uniformity	Glrlm	0.093679
	Long Run High Gray Level Emphasis	Glrlm	0.114855
	Zone Variance	Glszm	0.088184
	Large Area Emphasis	Glszm	0.084421
	Busyness	Ngtdm	0.110967
ALL-P	Maximum Probability	CMP-Glcm	0.141184
	Idn	CMP-Glcm	0.110634
	Low Gray Level Zone Emphasis	CMP-Glszm	0.105727
	Small Area Low Gray Level Emphasis	CMP-Glszm	0.113860
	Elongation	PP-Shape	0.101759
	Kurtosis	PP-First order	0.109194
	Small Area High Gray Level Emphasis	PP-Glszm	0.096117
	Maximum 3D Diameter	EP-Shape	0.083936
	Minimum	NCP-First order	0.137583
NCP：平扫期；CMP：皮质-髓质期；PP：实质期；EP：排泄期

下载: 导出CSV

表 3 测试集上5种模型鉴别肾癌亚型的性能

Table 3. Performance of five models in the test set in differentiating renal cell carcinoma subtypes

模型	准确率	精确率	敏感度	AUC(95%CI)	Kappa
平扫期模型	0.529	0.533	0.555	0.693(0.509~0.876)	0.163
皮髓质期模型	0.764	0.671	0.727	0.752(0.662~0.869)	0.446
实质期模型	0.783	0.689	0.734	0.811(0.659~0.912)	0.532
排泄期模型	0.711	0.62	0.665	0.725(0.575~0.831)	0.285
全期模型	0.810	0.717	0.799	0.853(0.768~0.889)	0.679
Kappa：0~1表示不同级别的一致性：0.00~0.20极低的一致性，0.21~0.40一般的一致性，0.41~0.60中等的一致性，0.61~0.80高度的一致性，0.81~1.00几乎完全一致.

下载: 导出CSV

参考文献(32)

[1]	Siegel RL, Miller KD, Jemal A. Cancer statistics, 2018[J]. CA Cancer J Clin, 2018, 68(1): 7-30. doi: 10.3322/caac.21442
[2]	Moch H, Cubilla AL, Humphrey PA, et al. The 2016 WHO classification of tumours of the urinary system and male genital organs- part A: renal, penile, and testicular tumours[J]. Eur Urol, 2016, 70(1): 93-105. doi: 10.1016/j.eururo.2016.02.029
[3]	Considine B, Hurwitz ME. Current status and future directions of immunotherapy in renal cell carcinoma[J]. Curr Oncol Rep, 2019, 21(4): 34. doi: 10.1007/s11912-019-0779-1
[4]	Barata PC, Rini BI. Treatment of renal cell carcinoma: Current status and future directions[J]. CA Cancer J Clin, 2017, 67(6): 507-24. doi: 10.3322/caac.21411
[5]	Robila V, Kraft AO, Smith SC. New entities, new technologies, new findings: a review of the cytologic features of recently established subtypes of renal cell carcinoma[J]. Cancer Cytopathol, 2019, 127 (2): 79-97. doi: 10.1002/cncy.22093
[6]	Berquist SW, Yim K, Ryan ST, et al. Systemic therapy in the management of localized and locally advanced renal cell carcinoma: Current state and future perspectives[J]. Int J Urol, 2019, 26(5): 532-42. doi: 10.1111/iju.13943
[7]	Ghali F, Patel SH, Derweesh IH. Current status of immunotherapy for localized and locally advanced renal cell carcinoma[J]. J Oncol, 2019, 2019: 7309205. http://www.ncbi.nlm.nih.gov/pubmed/31057615
[8]	Sasaguri K, Takahashi N. CT and MR imaging for solid renal mass characterization[J]. Eur J Radiol, 2018, 99: 40-54. doi: 10.1016/j.ejrad.2017.12.008
[9]	Young JR, Margolis D, Sauk S, et al. Clear cell renal cell carcinoma: discrimination from other renal cell carcinoma subtypes and oncocytoma at Multiphasic multidetector CT[J]. Radiology, 2013, 267(2): 444-53. doi: 10.1148/radiol.13112617
[10]	Yang CW, Shen SH, Chang YH, et al. Are there useful CT features to differentiate renal cell carcinoma from lipid-poor renal angiomyolipoma?[J]. Am J Roentgenol, 2013, 201(5): 1017-28. doi: 10.2214/AJR.12.10204
[11]	Young JR, Coy H, Kim HJ, et al. Performance of relative enhancement on Multiphasic MRI for the differentiation of clear cell renal cell carcinoma (RCC) from papillary and chromophobe RCC subtypes and oncocytoma[J]. AJR Am J Roentgenol, 2017, 208(4): 812-9. doi: 10.2214/AJR.16.17152
[12]	Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data[J]. Radiology, 2016, 278(2): 563-77. doi: 10.1148/radiol.2015151169
[13]	Liu ZY, Wang S, Dong D, et al. The applications of radiomics in precision diagnosis and treatment of oncology: opportunities and challenges[J]. Theranostics, 2019, 9(5): 1303-22. doi: 10.7150/thno.30309
[14]	Alessandrino F, Krajewski KM, Shinagare AB. Update on radiogenomics of clear cell renal cell carcinoma[J]. Eur Urol Focus, 2016, 2(6): 572-3. doi: 10.1016/j.euf.2017.01.012
[15]	Kocak B, Ates E, Durmaz ES, et al. Influence of segmentation margin on machine learning-based high-dimensional quantitative CT texture analysis: a reproducibility study on renal clear cell carcinomas[J]. Eur Radiol, 2019, 29(9): 4765-75. doi: 10.1007/s00330-019-6003-8
[16]	van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational radiomics system to decode the radiographic phenotype[J]. Cancer Res, 2017, 77(21): e104-e107. doi: 10.1158/0008-5472.CAN-17-0339
[17]	Dietterich TG. Ensemble methods in machine learning[M] //Multiple Classifier Systems. Berlin, Heidelberg: Springer Berlin Heidelberg, 2000: 1-15.
[18]	Bifet A, Holmes G, Pfahringer B, et al. New ensemble methods for evolving data streams[C] //Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining - KDD '09. June 28-July 1, 2009. Paris, France. New York: ACM Press, 2009(6): 139-48.
[19]	Reid S, Tibshirani R, Friedman J. A study of error variance estimation in Lasso regression[J]. Stat Sin, 2013, 26(1): 35-67. http://arxiv.org/abs/1311.5274
[20]	Ng F, Kozarski R, Ganeshan B, et al. Assessment of tumor heterogeneity by CT texture analysis: Can the largest crosssectional area be used as an alternative to whole tumor analysis? [J]. Eur J Radiol, 2013, 82(2): 342-8. doi: 10.1016/j.ejrad.2012.10.023
[21]	Cui EM, Lin F, Li Q, et al. Differentiation of renal angiomyolipoma without visible fat from renal cell carcinoma by machine learning based on whole-tumor computed tomography texture features[J]. Acta Radiol, 2019, 60(11): 1543-52. doi: 10.1177/0284185119830282
[22]	Varghese BA, Chen F, Hwang DH, et al. Differentiation of predominantly solid enhancing lipid-poor renal cell masses by use of contrast-enhanced CT: evaluating the role of texture in tumor subtyping[J]. AJR Am J Roentgenol, 2018, 211(6): W288-96. doi: 10.2214/AJR.18.19551
[23]	Zhou L, Zhang Z, Chen YC, et al. A deep learning-based radiomics model for differentiating benign and malignant renal tumors[J]. Transl Oncol, 2019, 12(2): 292-300. doi: 10.1016/j.tranon.2018.10.012
[24]	He XP, Zhang HM, Zhang T, et al. Predictive models composed by radiomic features extracted from multi-detector computed tomography images for predicting low- and high-grade clear cell renal cell carcinoma: a STARD- compliant article[J]. Medicine, 2019, 98(2): e13957. doi: 10.1097/MD.0000000000013957
[25]	Shu J, Tang Y, Cui J, et al. Clear cell renal cell carcinoma: CTbased radiomics features for the prediction of Fuhrman grade[J]. Eur J Radiol, 2018, 109: 8-12. doi: 10.1016/j.ejrad.2018.10.005
[26]	Ding J, Xing Z, Jiang Z, et al. CT- based radiomic model predicts high grade of clear cell renal cell carcinoma[J]. Eur J Radiol, 2018, 103: 51-6. doi: 10.1016/j.ejrad.2018.04.013
[27]	Yu H, Scalera J, Khalid M, et al. Texture analysis as a radiomic marker for differentiating renal tumors[J]. Abdom Radiol: NY, 2017, 42(10): 2470-8. doi: 10.1007/s00261-017-1144-1
[28]	Zhang GM, Shi B, Xue HD, et al. Can quantitative CT texture analysis be used to differentiate subtypes of renal cell carcinoma? [J]. Clin Radiol, 2019, 74(4): 287-94. doi: 10.1016/j.crad.2018.11.009
[29]	Kocak B, Yardimci AH, Bektas CT, et al. Textural differences between renal cell carcinoma subtypes: Machine learning- based quantitative computed tomography texture analysis with independent external validation[J]. Eur J Radiol, 2018, 107: 149-57. doi: 10.1016/j.ejrad.2018.08.014
[30]	Raman SP, Chen Y, Schroeder JL, et al. CT texture analysis of renal masses: pilot study using random forest classification for prediction of pathology[J]. Acad Radiol, 2014, 21(12): 1587-96. doi: 10.1016/j.acra.2014.07.023
[31]	Hodgdon T, McInnes MD, Schieda N, et al. Can quantitative CT texture analysis be used to differentiate fat- poor renal angiomyolipoma from renal cell carcinoma on unenhanced CT images?[J]. Radiology, 2015, 276(3): 787-96. doi: 10.1148/radiol.2015142215
[32]	Coy H, Hsieh K, Wu W, et al. Deep learning and radiomics: the utility of Google TensorFlow ^TM Inception in classifying clear cell renal cell carcinoma and oncocytoma on Multiphasic CT[J]. Abdom Radiol: NY, 2019, 44(6): 2009-20. doi: 10.1007/s00261-019-01929-0