en
×

分享给微信好友或者朋友圈

使用微信“扫一扫”功能。
通讯作者:

蒋健,E-mail:jiangjian72@sina.com

中图分类号:R739.41

文献标识码:A

文章编号:1007-4368(2023)05-634-06

DOI:10.7655/NYDXBNS20230507

参考文献 1
BRAUNSTEIN S,RALEIGH D,BINDRA R,et al.Pediat⁃ ric high ⁃ grade glioma:current molecular landscape and therapeutic approaches[J].J Neurooncol,2017,134(3):541-549
参考文献 2
万俊呈,陈洁.高级别脑胶质瘤患者复发因素及相关参数研究[J].系统医学,2021,6(21):174-177
参考文献 3
PENG W,CHEN S,KONG D,et al.Grade classification of human glioma using a convolutional neural network based on mid ⁃ infrared spectroscopy mapping[J].J Bio⁃ photonics,2022,15(4):e202100313
参考文献 4
MALTA T,DE SOUZA C,SABEDOT T,et al.Glioma CpG island methylator phenotype(G ⁃ CIMP):biological and clinical implications[J].Neuro Oncol,2018,20(5):608-620
参考文献 5
RUFF M,KIZILBASH S,BUCKNER J.Further under⁃ standing of glioma mechanisms of pathogenesis:implica⁃ tions for therapeutic development[J].Expert Rev Anti⁃ cancer Ther,2020,20(5):355-363
参考文献 6
THERESIA E,MALUEKA R,PRANACIPTA S,et al.As⁃ sociation between Ki⁃67 labeling index and histopatholog⁃ ical grading of glioma in indonesian population[J].Asian Pac J Cancer Prev,2020,21(4):1063-1068
参考文献 7
KLÖPPEL G,LA ROSA S.Correction to:Ki67 labeling in⁃ dex:assessment and prognostic role in gastroenteropancre⁃ atic neuroendocrine neoplasms[J].Virchows Arch,2018,472(3):515
参考文献 8
QU D,XU H,HAN X,et al.Expression of cyclinD1 and Ki⁃67 proteins in gliomas and its clinical significance[J].Eur Rev Med Pharmacol Sci,2014,18(4):516-519
参考文献 9
CAI J,YANG P,ZHANG C,et al.ATRX mRNA expres⁃ sion combined with IDH1/2 mutational status and Ki ⁃ 67 expression refines the molecular classification of astrocyt⁃ ic tumors:evidence from the whole transcriptome se⁃ quencing of 169 samples samples[J].Oncotarget,2014,5(9):2551-2561
参考文献 10
WANG P,LIU N,SONG H,et al.IDH⁃1R132H mutation status in diffuse glioma patients:implications for classifi⁃ cation[J].Oncotarget,2016,7(21):31393-31400
参考文献 11
CAMELO ⁃PIRAGUA S,KESARI S.Further understand⁃ ing of the pathology of glioma:implications for the clinic [J].Expert Rev Neurother,2016,16(9):1055-1065
参考文献 12
BARUA P D,CHAN W Y,DOGAN S,et al.Multilevel deep feature generation framework for automated detec⁃ tion of retinal abnormalities using OCT Images[J].Entro⁃ py(Basel),2021,23(12):1651
参考文献 13
WILLIAMS T,SAADAT L,GONEN M,et al.Radiomics in surgical oncology:applications and challenges[J].Comput Assist Surg(Abingdon,England),2021,26(1):85-96
参考文献 14
唐文天,张梓枫,尹建新,等.常规MRI特征在弥漫性星形细胞瘤IDH基因突变预测中的临床价值[J].南京医科大学学报(自然科学版),2022,42(3):376-381
参考文献 15
PINKER K,CHIN J,MELSAETHER A N,et al.Precision medicine and radiogenomics in breast cancer:new ap⁃ proaches toward diagnosis and treatment[J].Radiology,2018,287(3):732-747
参考文献 16
孙贤婷,廖伟华,曹冬,等.基于影像组学的logistic回归模型预测胶质瘤分级[J].中南大学学报(医学版),2021,46(4):385-392
参考文献 17
FU Q,LIU S L,HAO D P,et al.CT radiomics model for predicting the Ki⁃67 index of lung cancer:an exploratory study[J].Front Oncol,2021,11:743490
参考文献 18
ANGRAAL S,MORTAZAVI B,GUPTA A,et al.Machine learning prediction of mortality and hospitalization in heart failure with preserved ejection fraction[J].JACC Heart Failure,2020,8(1):12-21
参考文献 19
BILIMORIA K,LIU Y,PARUCH J,et al.Development and evaluation of the universal ACS NSQIP surgical risk calculator:a decision aid and informed consent tool for patients and surgeons[J].J Am Coll Surg,2013,217(5):833-842
参考文献 20
MA H,BANDOS A,GUR D.On the use of partial area un⁃ der the ROC curve for comparison of two diagnostic tests [J].Biom J,2015,57(2):304-320
参考文献 21
VAN CALSTER B,NIEBOER D,et al.A calibration hier⁃ archy for risk models was defined:from utopia to empiri⁃ cal data[J].J Clin Epidemiol,2016,74:167-176
参考文献 22
TOM M C,CAHILL D P,BUCKNER J C,et al.Manage⁃ ment for different glioma subtypes:are all low⁃grade glio⁃ mas created equal?[J].Am Soc Clin Oncol Educ Book,2019,39:133-145
参考文献 23
MORGAN L.The epidemiology of glioma in adults:a “state of the science”review[J].Neuro Oncol,2015,17(4):623-624
参考文献 24
DEBOM G,RUBENICH D,BRAGANHOL E.Adenosin⁃ ergic signaling as a key modulator of the glioma microen⁃ vironment and reactive astrocytes[J].Front Neurosci,2021,15:648476
参考文献 25
RADIN D P,TSIRKA S E.Interactions between tumor cells,neurons,and microglia in the glioma microenviron⁃ ment[J].Int J Mol Sci,2020,21(22):8476
参考文献 26
SANTISTEBAN M,REYNOLDS C,BARR FRITCHER E,et al.Ki67:a time⁃varying biomarker of risk of breast can⁃ cer in atypical hyperplasia[J].Breast Cancer Res Treat,2010,121(2):431-437
参考文献 27
HOSNY A,PARMAR C,QUACKENBUSH J,et al.Artifi⁃ cial intelligence in radiology[J].Nat Rev Cancer,2018,18(8):500-510
参考文献 28
SAH B,OWCZARCZYK K,SIDDIQUE M,et al.Ra⁃ diomics in esophageal and gastric cancer[J].Abdom Ra⁃ diol(NY),2019,44(6):2048-2058
参考文献 29
GILLIES R,SCHABATH M.Radiomics improves cancer screening and early detection[J].Cancer Epidemiol Bio⁃ markers Prev,2020,29(12):2556-2567
参考文献 30
徐青青,单文莉,朱艳,等.基于CT影像组学对孤立性肺结节性质分类的预测效能[J].南京医科大学学报(自然科学版),2021,41(4):617-623
参考文献 31
MUN S,WONG K,LO S,et al.Artificial intelligence for the future radiology diagnostic service[J].Front Mol Bios⁃ ci,2020,7:614258
参考文献 32
KERNBACH J M,ORT J,HAKVOORT K,et al.Introduc⁃ tion to machine learning in neuroimaging[J].Acta Neuro⁃ chir Suppl,2022,134:121-124
参考文献 33
ENG J.Receiver operating characteristic analysis:a prim⁃ er[J].Acad Radiol,2005,12(7):909-916
参考文献 34
SUBTIL F,RABILLOUD M.An enhancement of ROC curves made them clinically relevant for diagnostic ⁃ test comparison and optimal ⁃ threshold determination[J].J Clin Epidemiol,2015,68(7):752-759
目录contents

    摘要

    目的:基于胶质瘤增强T1加权成像(enhancement T1-weighted image,T1WI+C)影像组学特征进行机器学习,构建预测胶质瘤细胞增殖活性的Ki67指数预测模型。方法:回顾性分析本院手术病理结果为胶质瘤并经免疫组化测定Ki67指数的患者113例,随机按约8∶2拆分为训练集与测试集。通过MRIcroGL软件手绘胶质瘤感兴趣区域(region of interest,ROI),利用 Python中 pyradiomics模块获得1338个影像特征,通过t检验与最小绝对收缩和选择算子(least absolute shrinkage and selection operator,Lasso)回归算法筛选最佳影像组学特征,基于最佳特征使用支持向量机(support vector machine,SVM)分类器建立 Ki67预测模型,利用受试者操作特征曲线(receiver operating characteristic curve,ROC曲线)与模型校准曲线评估模型的预测效能。结果:每例患者T1WI+C图像提取1338个影像组学特征,降维后筛选出 6个与胶质瘤Ki67指数密切相关的特征。基于支持向量机算法模型在训练集Ki67指数预测中曲线下面积(area under the curve,AUC)为0.82、准确率为0.72;在测试集Ki67指数预测中AUC为0.91、准确率为0.83。模型校准曲线结果显示布尔里得分为0.175。结论:基于T1WI+C的影像建立支持向量机预测模型对胶质瘤细胞增殖活性可能具有较好的预测效能,有助于患者个体化诊疗和未来精准化医疗的发展。

    Abstract

    Objective:The current machine learning was performed based on the radiomic features of contrast - enhanced T1 - weighted imaging(T1WI+C)of glioma,which was used to construct a prediction model of Ki67 index for predicting the proliferation activity of glioma cells. Methods:A retrospective analysis onsisted 113 patients with glioma,which were confirmed by the surgical and pathological results of our hospital and the Ki67 index defined by immunohistochemistry. The training set and the test set were divided into 8∶2. The glial region of interest(ROI)was hand-painted by MRIcroGL software,1338 image features were obtained by using the pyradiomics module in Python,and the best image features were obtained through t-test and the least absolute shrinkage and selection operator(Lasso)regression algorithm. Furthermore,using the support vector machine classifier to build the Ki67 prediction model based on the best features,using the receiver operating characteristic(ROC)curve and the calibration curve to evaluate the predictive performance of the model. Results:A total of 1338 radiomics features were extracted from T1WI+C images of each patient,and six features closely related to glioma ki67 index were screened out after dimensionality reduction. Based on the support vector machine algorithm model,the AUC in the training set of Ki67 index prediction was 0.82,and the accuracy rate was 0.72;in the test set of Ki67 index prediction,the AUC was 0.91,and the accuracy rate was 0.83. The results of the calibration curve of the model showed that the difference was not statistically significant(Brier score=0.175). Conclusion:The prediction model of support vector machine that established on the learning of the T1W1+C imaging omics characteristics may have a better predictive effert on the proliferation activity of glioma cells,and it may help the selection of individual diagnosis and treatment plans for patients and the development of precision medical care in the future.

    关键词

    磁共振图像胶质瘤Ki67

    Keywords

    MRIgliomasKi67

  • 胶质瘤是中枢神经系统最常见肿瘤[1]。我国脑胶质瘤年发病率为5~8/10万,5年病死率在全身肿瘤中仅次于胰腺癌和肺癌[2]。临床实践中,胶质瘤治疗方案与预后常借助病理与免疫组化共同决定[3]。随着胶质瘤发生机制的深入研究,更多有助于临床诊疗和判断预后的分子标志物被相继发现。肿瘤细胞的增殖方式和增殖速度与胶质瘤的恶性程度及预后有极大关联[4-5]。Ki67作为细胞增殖阶段表达的一种核抗原,很大程度上反映细胞的增殖活性[6-7]。正常脑组织中Ki67表达水平较低,胶质瘤细胞中Ki67 表达水平明显较正常脑组织高,同时随着胶质瘤病理等级提高,Ki67表达水平也显著提高,低级别胶质瘤 Ki67 显著低于高级别胶质瘤[8]。 Ki67 存在半衰期较短的特性,在评价肿瘤的增殖活性方面优于半衰期较长的增殖细胞核抗原。相关报道也表明Ki67与突变的异柠檬酸脱氢酶1(IDH1)基因密切相关,因此 Ki67 可以作为胶质瘤的预后指标[9-10]。目前Ki67指数还是依靠病理和免疫组化获得,必须通过活检和手术,缺乏一种术前无创预测 Ki67指数的方法[11]。随着影像组学、人工智能的发展,医疗诊断及治疗方案已较前不断更新进步[12-13]。基于临床医学影像特征的多维度研究,不仅可预测肿瘤基因突变类型,同时对临床制定治疗方案和评估预后具有重要帮助[14]。影像组学通过提取疾病影像数据中包含的定量特征,找出疾病的影像标志物,具有预测疾病分类、分级、无创、经济、高效和可重复及有助于临床决策等优点[15-16]。本研究尝试通过影像组学与人工智能结合[17],以无创、方便、经济方式预测胶质瘤 Ki67 表达水平,以期满足临床胶质瘤术前增殖活性程度预测及个性化治疗需求。

  • 1 对象和方法

  • 1.1 对象

  • 本研究回顾收集南京医科大学附属淮安第一医院经手术病理结果证实的胶质瘤患者334例,剔除没有免疫组化结果者102例,剔除没有MRI增强序列者113例,剔除同一诊断时期患有其他肿瘤患者病 5 例,最终获取 113 例作为本研究对象。本研究经南京医科大学附属淮安第一医院伦理委员会批准,免除受试者知情同意。

  • 1.2 方法

  • 1.2.1 MRI扫描

  • 采用 SIEMENS Aera1.5T 磁共振成像仪及20通道头部相控阵线圈采集患者的颅脑 MRI 增强数据。轴位增强 T1加权成像(enhancement T1⁃weight⁃ ed image,T1WI+C)序列参数:重复时间(time of rep⁃ etition,TR)350 ms,回波时间(time of echo,TE) 89 ms,矩阵320×224,层厚5 mm,层间距1.5 mm,层数 20,视野(field of view,FOV)230 mm×230 mm。

  • 1.2.2 图像分析

  • 图像来自南京医科大学附属淮安第一医院图像归档和通讯系统(picture archiving and communi⁃ cation systems,PACS)。考虑原始图像之间参数存在一定差异性,选择在提取胶质瘤感兴趣区(region of interest,ROI)前优先对原序列图像素进预处理。预处理包含重采样、偏置场校正:图像重采统一使患者MRI图像的灰度值分布在1 k~2 k,考虑图像对比度或像素亮度不一致对后续图像特征采集存在一定偏差,将不同设备或扫描参数差异进行最小化,以0.5×0.5×0.5体素进行重采样。使用MRIcroGL图像软件手绘ROI。患者有效的3D ROI主要由2名具有10年以上工作经验的神经外科医师独立手动绘制,然后提交给另一位具有20年工作经验的神经外科高级医师进行审核,若审核医师评定结果存在较大偏倚,则对原3D ROI进行重新修改或剔除。最终使用的 ROI 为 2 名医师各自勾画 ROI 最有效区。 ROI勾画标准为肿瘤特异性导致的强化区域,排除肿瘤周边水肿区。

  • 1.2.3 特征提取

  • 影像特征提取基于 Python3.7 的 pyradiomics 模块实现,最终通过 First Order、Shape、Glcm、Glszm、 Glrlm、Ngtdm、Glim 等获得 1 338 个影像相关特征,特征定义和计算机算法在 pyradiomics:https//pyradiomics.readthedocs.io/en/latest/官网可获得。

  • 1.2.4 影像特征筛选与降维

  • 通过 pyradiomics 模块获取的影像特征并非都与患者病理结果Ki67具有相关性,同时考虑提取数据变量存在差异性,先将经病理结果证实的Ki67≤ 10设立影像组学标签“0”,将病理证实的Ki67>10设立影像组学标签“1”,然后将所有数据集进行归一化处理,再将标准化的数据进行 t 检验剔除两组数据间无统计学意义特征,最后通过最小绝对收缩与选择算子Lasso算法,同时进行10次交叉验证,最终引入最优惩罚系数λ,对沉余量进行惩罚,逐步将不重要参数压缩至0,降低自变量维度,从而进一步筛选出具有显著意义的影像特征。

  • 1.2.5 影像组学模型训练与验证

  • 影像组学联合机器学习模型的建立基于Python (3.7.0)和机器学习库 sklearn等导入实现,通过将数据集以8∶2拆分为训练集与测试集。将筛选后的数据进行交叉验证,网格搜索获取最优C值与gamma 值等优化参数,再通过支持向量机学习算法进行训练并建立相应的影像组学模型,测试集数据用于对训练好的模型进行验证,最后结果以准确度、灵敏度、特异度和受试者操作特征(receiver operat char⁃ acterist,ROC)曲线下面积(area under the curve, AUC)表示。

  • 1.3 统计学方法

  • 本研究采样 SPSS 软件进行相关统计学分析。使用卡方检验和独立样本t检验分析胶质瘤患者高表达Ki67与低表达Ki67组年龄、性别、是否合并高血压和是否诱发癫痫组间差异是否具有统计学意义,P<0.05 为差异具有统计学意义。采样组间相关系数(inter classcorrelation coefficient,ICC)评估 2 名神经医师各自获取病灶ROL影像特征的一致性, ICC>0.75 被认为一致性良好[16]。用 ROC 曲线与模型校准曲线评估模型区分效能,评估指标为AUC与布里尔得分(Brier score)[18-19],分数越高预测结果越差,分数越低模型预测能力越好。

  • 2 结果

  • 2.1 患者基本临床信息与Ki67指数的关系

  • 本研究共纳入的113例胶质瘤患者中Ki67≤10 共 50 例(男 27 例,女 23 例),平均年龄(50.7±13.7) 岁,Ki67>10共63例(男36例,女27例),平均年龄 (52.9 ± 13.0)岁。结果发现胶质瘤 Ki67≤10 和 Ki67>10之间的年龄、性别、高血压史、癫痫史差异没有统计学意义(P>0.05,表1)。

  • 表1 患者基本临床信息与Ki67指数的关系

  • Table1 Relation between patient basic clinical informa⁃ tion and the Ki67 index

  • 2.2 特征筛选结果

  • 2 名神经外科医师提取的 ROI 影像特征 ICC>0.75,表明 2 名医师提取 ROI 具有较高的组间一致性。将所有数据集通过t检验共筛选出608个具有差异性影像特征,选择Lasso回归,利用10次交叉验证中最优的λ选择系数非零的最佳影像组学特征 (图1A),根据影像组学特征的 Lasso 系数分布(图1B),绘制最佳特征与权重图(图1C),同时将Lasso 回归筛选出6个特征进行组间差异分析并绘制热度图(图1D)。

  • 2.3 模型训练与预测结果

  • 通过交叉网格搜索得到 C 为 0.5、gamma 为 0.062 5,将最佳影像组学特征通过SVM算法建立胶质瘤Ki67级别预测模型,应用 ROC 曲线分析模型,在训练集中的AUC 值、灵敏度、特异度、准确度分别为0.82、0.71、0.79、0.72,在测试集中AUC 值、灵敏度、特异度、准确度分别为0.91、0.88、0.88、0.83(图2)。

  • 2.4 建立模型校准曲线

  • 为进一步验证预测概率与实际概率之间是否较大偏倚,利用预测模型校准曲线对结果进行分析[20-21]。图3可示实线代表的校准线与理想模型代表的对角虚线存在一定偏差,但较接近吻合,布里尔得分0.175。

  • 3 讨论

  • 胶质瘤是中枢神经系统占比最多、恶性程度最高、预后最差的肿瘤[22-23],患者常存在明显的异质性,手术联合术后放化疗仍是主要治疗策略。胶质瘤组织中存在大量不同类型细胞,这些细胞共同组成胶质瘤增殖侵袭的特殊微环境[24-25]。随着肿瘤相关免疫组化的发展,越来越多的肿瘤相关标志物被相继发现,如 Ki67、IDH、ATRX、BRAT、TP53 等,为胶质瘤的诊疗带来新的突破。其中反映细胞增殖活性及肿瘤恶性程度的细胞增殖标志物Ki67指数常作为脑胶质瘤恶性程度和预后的评估指标之一[26],但 Ki67指数只能通过术后病理免疫组化获得,亟待一种无创的方式在术前大致掌握肿瘤Ki67指数水平,评估胶质瘤恶性程度,以制定更准确的手术策略。近年来,影像组学越来越成为研究热点[27]。影像组学是以影像特征为基础的一种辅助临床决策的新兴研究策略[28]。目前广泛应用于肿瘤良恶性鉴别、术前肿瘤分期预测、基因突变及生存预后等领域[29-30]。相比于传统医学影像资料分析,基于人工智能下的图像分析具有更大能力,运用于影像组学中也具有更大潜能[31]。大数据背景下医学影像与人工智能的多学科交叉,可能为未来精准医疗的发展带来新的契机[32]

  • 图1 Lasso回归运用于特征筛选

  • Figure1 Feature selection using Lasso

  • 图2 模型预测训练集(A)和测试集(B)在胶质瘤Ki67指数水平的受试者操作特征曲线

  • Figure2 Receiver operating characteristic curve of training set(A)and test set(B)for predicting the level of Ki67 indexes of gliomas

  • 图3 预测模型校准曲线

  • Figure3 Calibration plot for the predictive model

  • 本研究通过算法提取高通量特征,深入挖掘影像图像中潜在的临床价值,同时选择对高维度数据分析算法 Lasso 回归模型进行最终特征提取,同时为避免特征出现过拟合现象,利用交叉验证选择最优参数,最后共筛选出6项最有价值影像特征。基于术前影像组学筛选的最优特征与SVM机器学习模型作为分类器,对胶质瘤细胞增殖活性进行预测,该研究采用ROC曲线的AUC值等对模型训练集和测试集进行评估,同时采用校准曲线的布尔里得分评估模型对胶质瘤 Ki67 指数分类的实际偏倚。 AUC 越接近 1,表明分类效能越好[33-34]。本研究构建的模型在训练集、测试集AUC分别为0.82、0.91,表明预测模型具有较好的区分能力;同时校准曲线与理想曲线偏差小,吻合度较好(布尔里得分为 0.175),表明该模型具有较好的预测效能,有望成为术前预测胶质瘤细胞增殖活性的辅助方法。

  • 本研究影像特征提取通过 ROI 勾画获得,ROI 勾画方式很多,有手动勾画、半自动勾画、自动化勾画等。大多认为计算机自动算法勾画 ROI 省时省力,同时避免源于主观因素影响。也有学者认为影像科专业医师基于专业背景更有助于绘制准确的ROI。但本研究基于影像图像及病理状态,图像特征与肿瘤术中形态、质地等具有很大差异性,由经验丰富的神经外科医师绘制可能更具有效性。

  • 本研究的局限性在于:①研究样本量相对不足,因此仅对训练模型进行了内部测试,未进行外部验证,后续会进一步引入多中心外部验证;②本研究仅通过影像组学特征建立单一Ki67预测模型,在后续研究中会引入更多临床资料及多种相关免疫组化预测模型进一步完善不足;③本研究采用支持向量机建模,而机器学习模型种类多,如逻辑分析、决策树、随机森林、梯度提升树等,在后续的研究中我们会增加多中心研究进行多模型比较,同时进一步优化模型,弥补不足,以期为临床应用提供更优的模型。

  • 综上所述,基于 T1WI+C的支持向量机算法构建的胶质瘤细胞增殖活性预测模型,可能是术前预测细胞增殖活性程度的一种新途径,有望为患者个性化治疗提供参考依据。

  • 参考文献

    • [1] BRAUNSTEIN S,RALEIGH D,BINDRA R,et al.Pediat⁃ ric high ⁃ grade glioma:current molecular landscape and therapeutic approaches[J].J Neurooncol,2017,134(3):541-549

    • [2] 万俊呈,陈洁.高级别脑胶质瘤患者复发因素及相关参数研究[J].系统医学,2021,6(21):174-177

    • [3] PENG W,CHEN S,KONG D,et al.Grade classification of human glioma using a convolutional neural network based on mid ⁃ infrared spectroscopy mapping[J].J Bio⁃ photonics,2022,15(4):e202100313

    • [4] MALTA T,DE SOUZA C,SABEDOT T,et al.Glioma CpG island methylator phenotype(G ⁃ CIMP):biological and clinical implications[J].Neuro Oncol,2018,20(5):608-620

    • [5] RUFF M,KIZILBASH S,BUCKNER J.Further under⁃ standing of glioma mechanisms of pathogenesis:implica⁃ tions for therapeutic development[J].Expert Rev Anti⁃ cancer Ther,2020,20(5):355-363

    • [6] THERESIA E,MALUEKA R,PRANACIPTA S,et al.As⁃ sociation between Ki⁃67 labeling index and histopatholog⁃ ical grading of glioma in indonesian population[J].Asian Pac J Cancer Prev,2020,21(4):1063-1068

    • [7] KLÖPPEL G,LA ROSA S.Correction to:Ki67 labeling in⁃ dex:assessment and prognostic role in gastroenteropancre⁃ atic neuroendocrine neoplasms[J].Virchows Arch,2018,472(3):515

    • [8] QU D,XU H,HAN X,et al.Expression of cyclinD1 and Ki⁃67 proteins in gliomas and its clinical significance[J].Eur Rev Med Pharmacol Sci,2014,18(4):516-519

    • [9] CAI J,YANG P,ZHANG C,et al.ATRX mRNA expres⁃ sion combined with IDH1/2 mutational status and Ki ⁃ 67 expression refines the molecular classification of astrocyt⁃ ic tumors:evidence from the whole transcriptome se⁃ quencing of 169 samples samples[J].Oncotarget,2014,5(9):2551-2561

    • [10] WANG P,LIU N,SONG H,et al.IDH⁃1R132H mutation status in diffuse glioma patients:implications for classifi⁃ cation[J].Oncotarget,2016,7(21):31393-31400

    • [11] CAMELO ⁃PIRAGUA S,KESARI S.Further understand⁃ ing of the pathology of glioma:implications for the clinic [J].Expert Rev Neurother,2016,16(9):1055-1065

    • [12] BARUA P D,CHAN W Y,DOGAN S,et al.Multilevel deep feature generation framework for automated detec⁃ tion of retinal abnormalities using OCT Images[J].Entro⁃ py(Basel),2021,23(12):1651

    • [13] WILLIAMS T,SAADAT L,GONEN M,et al.Radiomics in surgical oncology:applications and challenges[J].Comput Assist Surg(Abingdon,England),2021,26(1):85-96

    • [14] 唐文天,张梓枫,尹建新,等.常规MRI特征在弥漫性星形细胞瘤IDH基因突变预测中的临床价值[J].南京医科大学学报(自然科学版),2022,42(3):376-381

    • [15] PINKER K,CHIN J,MELSAETHER A N,et al.Precision medicine and radiogenomics in breast cancer:new ap⁃ proaches toward diagnosis and treatment[J].Radiology,2018,287(3):732-747

    • [16] 孙贤婷,廖伟华,曹冬,等.基于影像组学的logistic回归模型预测胶质瘤分级[J].中南大学学报(医学版),2021,46(4):385-392

    • [17] FU Q,LIU S L,HAO D P,et al.CT radiomics model for predicting the Ki⁃67 index of lung cancer:an exploratory study[J].Front Oncol,2021,11:743490

    • [18] ANGRAAL S,MORTAZAVI B,GUPTA A,et al.Machine learning prediction of mortality and hospitalization in heart failure with preserved ejection fraction[J].JACC Heart Failure,2020,8(1):12-21

    • [19] BILIMORIA K,LIU Y,PARUCH J,et al.Development and evaluation of the universal ACS NSQIP surgical risk calculator:a decision aid and informed consent tool for patients and surgeons[J].J Am Coll Surg,2013,217(5):833-842

    • [20] MA H,BANDOS A,GUR D.On the use of partial area un⁃ der the ROC curve for comparison of two diagnostic tests [J].Biom J,2015,57(2):304-320

    • [21] VAN CALSTER B,NIEBOER D,et al.A calibration hier⁃ archy for risk models was defined:from utopia to empiri⁃ cal data[J].J Clin Epidemiol,2016,74:167-176

    • [22] TOM M C,CAHILL D P,BUCKNER J C,et al.Manage⁃ ment for different glioma subtypes:are all low⁃grade glio⁃ mas created equal?[J].Am Soc Clin Oncol Educ Book,2019,39:133-145

    • [23] MORGAN L.The epidemiology of glioma in adults:a “state of the science”review[J].Neuro Oncol,2015,17(4):623-624

    • [24] DEBOM G,RUBENICH D,BRAGANHOL E.Adenosin⁃ ergic signaling as a key modulator of the glioma microen⁃ vironment and reactive astrocytes[J].Front Neurosci,2021,15:648476

    • [25] RADIN D P,TSIRKA S E.Interactions between tumor cells,neurons,and microglia in the glioma microenviron⁃ ment[J].Int J Mol Sci,2020,21(22):8476

    • [26] SANTISTEBAN M,REYNOLDS C,BARR FRITCHER E,et al.Ki67:a time⁃varying biomarker of risk of breast can⁃ cer in atypical hyperplasia[J].Breast Cancer Res Treat,2010,121(2):431-437

    • [27] HOSNY A,PARMAR C,QUACKENBUSH J,et al.Artifi⁃ cial intelligence in radiology[J].Nat Rev Cancer,2018,18(8):500-510

    • [28] SAH B,OWCZARCZYK K,SIDDIQUE M,et al.Ra⁃ diomics in esophageal and gastric cancer[J].Abdom Ra⁃ diol(NY),2019,44(6):2048-2058

    • [29] GILLIES R,SCHABATH M.Radiomics improves cancer screening and early detection[J].Cancer Epidemiol Bio⁃ markers Prev,2020,29(12):2556-2567

    • [30] 徐青青,单文莉,朱艳,等.基于CT影像组学对孤立性肺结节性质分类的预测效能[J].南京医科大学学报(自然科学版),2021,41(4):617-623

    • [31] MUN S,WONG K,LO S,et al.Artificial intelligence for the future radiology diagnostic service[J].Front Mol Bios⁃ ci,2020,7:614258

    • [32] KERNBACH J M,ORT J,HAKVOORT K,et al.Introduc⁃ tion to machine learning in neuroimaging[J].Acta Neuro⁃ chir Suppl,2022,134:121-124

    • [33] ENG J.Receiver operating characteristic analysis:a prim⁃ er[J].Acad Radiol,2005,12(7):909-916

    • [34] SUBTIL F,RABILLOUD M.An enhancement of ROC curves made them clinically relevant for diagnostic ⁃ test comparison and optimal ⁃ threshold determination[J].J Clin Epidemiol,2015,68(7):752-759

  • 参考文献

    • [1] BRAUNSTEIN S,RALEIGH D,BINDRA R,et al.Pediat⁃ ric high ⁃ grade glioma:current molecular landscape and therapeutic approaches[J].J Neurooncol,2017,134(3):541-549

    • [2] 万俊呈,陈洁.高级别脑胶质瘤患者复发因素及相关参数研究[J].系统医学,2021,6(21):174-177

    • [3] PENG W,CHEN S,KONG D,et al.Grade classification of human glioma using a convolutional neural network based on mid ⁃ infrared spectroscopy mapping[J].J Bio⁃ photonics,2022,15(4):e202100313

    • [4] MALTA T,DE SOUZA C,SABEDOT T,et al.Glioma CpG island methylator phenotype(G ⁃ CIMP):biological and clinical implications[J].Neuro Oncol,2018,20(5):608-620

    • [5] RUFF M,KIZILBASH S,BUCKNER J.Further under⁃ standing of glioma mechanisms of pathogenesis:implica⁃ tions for therapeutic development[J].Expert Rev Anti⁃ cancer Ther,2020,20(5):355-363

    • [6] THERESIA E,MALUEKA R,PRANACIPTA S,et al.As⁃ sociation between Ki⁃67 labeling index and histopatholog⁃ ical grading of glioma in indonesian population[J].Asian Pac J Cancer Prev,2020,21(4):1063-1068

    • [7] KLÖPPEL G,LA ROSA S.Correction to:Ki67 labeling in⁃ dex:assessment and prognostic role in gastroenteropancre⁃ atic neuroendocrine neoplasms[J].Virchows Arch,2018,472(3):515

    • [8] QU D,XU H,HAN X,et al.Expression of cyclinD1 and Ki⁃67 proteins in gliomas and its clinical significance[J].Eur Rev Med Pharmacol Sci,2014,18(4):516-519

    • [9] CAI J,YANG P,ZHANG C,et al.ATRX mRNA expres⁃ sion combined with IDH1/2 mutational status and Ki ⁃ 67 expression refines the molecular classification of astrocyt⁃ ic tumors:evidence from the whole transcriptome se⁃ quencing of 169 samples samples[J].Oncotarget,2014,5(9):2551-2561

    • [10] WANG P,LIU N,SONG H,et al.IDH⁃1R132H mutation status in diffuse glioma patients:implications for classifi⁃ cation[J].Oncotarget,2016,7(21):31393-31400

    • [11] CAMELO ⁃PIRAGUA S,KESARI S.Further understand⁃ ing of the pathology of glioma:implications for the clinic [J].Expert Rev Neurother,2016,16(9):1055-1065

    • [12] BARUA P D,CHAN W Y,DOGAN S,et al.Multilevel deep feature generation framework for automated detec⁃ tion of retinal abnormalities using OCT Images[J].Entro⁃ py(Basel),2021,23(12):1651

    • [13] WILLIAMS T,SAADAT L,GONEN M,et al.Radiomics in surgical oncology:applications and challenges[J].Comput Assist Surg(Abingdon,England),2021,26(1):85-96

    • [14] 唐文天,张梓枫,尹建新,等.常规MRI特征在弥漫性星形细胞瘤IDH基因突变预测中的临床价值[J].南京医科大学学报(自然科学版),2022,42(3):376-381

    • [15] PINKER K,CHIN J,MELSAETHER A N,et al.Precision medicine and radiogenomics in breast cancer:new ap⁃ proaches toward diagnosis and treatment[J].Radiology,2018,287(3):732-747

    • [16] 孙贤婷,廖伟华,曹冬,等.基于影像组学的logistic回归模型预测胶质瘤分级[J].中南大学学报(医学版),2021,46(4):385-392

    • [17] FU Q,LIU S L,HAO D P,et al.CT radiomics model for predicting the Ki⁃67 index of lung cancer:an exploratory study[J].Front Oncol,2021,11:743490

    • [18] ANGRAAL S,MORTAZAVI B,GUPTA A,et al.Machine learning prediction of mortality and hospitalization in heart failure with preserved ejection fraction[J].JACC Heart Failure,2020,8(1):12-21

    • [19] BILIMORIA K,LIU Y,PARUCH J,et al.Development and evaluation of the universal ACS NSQIP surgical risk calculator:a decision aid and informed consent tool for patients and surgeons[J].J Am Coll Surg,2013,217(5):833-842

    • [20] MA H,BANDOS A,GUR D.On the use of partial area un⁃ der the ROC curve for comparison of two diagnostic tests [J].Biom J,2015,57(2):304-320

    • [21] VAN CALSTER B,NIEBOER D,et al.A calibration hier⁃ archy for risk models was defined:from utopia to empiri⁃ cal data[J].J Clin Epidemiol,2016,74:167-176

    • [22] TOM M C,CAHILL D P,BUCKNER J C,et al.Manage⁃ ment for different glioma subtypes:are all low⁃grade glio⁃ mas created equal?[J].Am Soc Clin Oncol Educ Book,2019,39:133-145

    • [23] MORGAN L.The epidemiology of glioma in adults:a “state of the science”review[J].Neuro Oncol,2015,17(4):623-624

    • [24] DEBOM G,RUBENICH D,BRAGANHOL E.Adenosin⁃ ergic signaling as a key modulator of the glioma microen⁃ vironment and reactive astrocytes[J].Front Neurosci,2021,15:648476

    • [25] RADIN D P,TSIRKA S E.Interactions between tumor cells,neurons,and microglia in the glioma microenviron⁃ ment[J].Int J Mol Sci,2020,21(22):8476

    • [26] SANTISTEBAN M,REYNOLDS C,BARR FRITCHER E,et al.Ki67:a time⁃varying biomarker of risk of breast can⁃ cer in atypical hyperplasia[J].Breast Cancer Res Treat,2010,121(2):431-437

    • [27] HOSNY A,PARMAR C,QUACKENBUSH J,et al.Artifi⁃ cial intelligence in radiology[J].Nat Rev Cancer,2018,18(8):500-510

    • [28] SAH B,OWCZARCZYK K,SIDDIQUE M,et al.Ra⁃ diomics in esophageal and gastric cancer[J].Abdom Ra⁃ diol(NY),2019,44(6):2048-2058

    • [29] GILLIES R,SCHABATH M.Radiomics improves cancer screening and early detection[J].Cancer Epidemiol Bio⁃ markers Prev,2020,29(12):2556-2567

    • [30] 徐青青,单文莉,朱艳,等.基于CT影像组学对孤立性肺结节性质分类的预测效能[J].南京医科大学学报(自然科学版),2021,41(4):617-623

    • [31] MUN S,WONG K,LO S,et al.Artificial intelligence for the future radiology diagnostic service[J].Front Mol Bios⁃ ci,2020,7:614258

    • [32] KERNBACH J M,ORT J,HAKVOORT K,et al.Introduc⁃ tion to machine learning in neuroimaging[J].Acta Neuro⁃ chir Suppl,2022,134:121-124

    • [33] ENG J.Receiver operating characteristic analysis:a prim⁃ er[J].Acad Radiol,2005,12(7):909-916

    • [34] SUBTIL F,RABILLOUD M.An enhancement of ROC curves made them clinically relevant for diagnostic ⁃ test comparison and optimal ⁃ threshold determination[J].J Clin Epidemiol,2015,68(7):752-759