基于多参数影像组学术前预测原发性中枢神经系统淋巴瘤Ki-67表达的价值

打开文本图片集

Preoperativepredictionof multiparametricradiomics forKi-67expressionlevel inprimarycentral nervoussystem .ymphoma

LIXie，LIHuihu’，LIFan，WANYun²

'Departmentofuclearedicine，omingeople'sospital，oming4ina;Departmentofdiology，iniple's Hospital，Xinyi525300，China.

[Abstract]Objective：Toexplorethevalueofconstructingamachinelearning modelbasedonmultiparametricradiomics forpreoperativeidentificationofKi-67expresionlevelinprimarycentralnervoussystemlymphoma（PCNSL）．Methods：A retrospectiveanalysiswasconstructedon203PCNSLpatients，with162patientsfromonecenterservingasthetraining setand41patientsfromanothercenterastheexternalvalidationset.Manualsegmentationof lesionsonTWI-CE，DWI， and T2 FLAIRimageswasperformed to extract theoptimal radiomics features.Four machine learningalgorithms wereused toconstructradiomicmodelsbasedontheoptimalradiomicsfeatures.Independentclinicalpredictorswereidentifidthrough univariateandmultivariatelogisticregressionanalysis.Acombinedmodelincorporatingthesepredictorsandtheradiomics scorebasedontheoptimalradiomicmodelwasdevelopedtopredictKi-67expresionlevel，andthepredictiveperformance ofthecombinedmodelwasevaluated.Results：Therewere203cases，including89caseswithKi-67highexpresionand 114 cases with low expression.Tumor maximum diameter，peritumoral edema，maximum standardized uptake value（ SUVmax ） on 18F -FDGPET/CTwereindependentpredictorsofKi-67expresionlevel.Utimately，19radiomicsfeatureswerescreened out，and fourmachinelearningalgorithmswereusedtoconstructtheradiomicsmodel，among which，thegradientboosting machine（GBM）modelhadahighAUCvalue（O.88forthetrainingsetandO.83fortheexternalvalidationset）.The RandomForest（RF）modelfolowed，withanAUCvalueofO.86forthetrainingsetandO.79fortheexternalvalidationset. TheGBMmodelhadhighersensitivityandrecallrate，whiletheRFmodelhadahigheraccuracy.ThecombinedGBMmodel hadthebestpredictiveperformance，withthehighestAUCvaluesforthetrainingsetandexternalvalidationsetbeing 0.92（95%CI0.88-0.96）and0.88（95%CI0.78-0.97），respectively.Calibrationcurvesindicatedgoodcalibrationofthe combinedGBM model.Decisioncurveanalysisshowed thatthecombinedGBMmodelhadahigheroverallnet benefit. Conclusion：AmultiparametricradiomicsmodelcombinedwithclinicalfeaturescanefectivelystratifythepreoperativeKi-67 expression leveL in PCNSL.

Keywords]Machine learning;Radiomics;Primary central nervoussystem lymphoma;Ki-67;Magneticresonance imaging

原发性中枢神经系统淋巴瘤（primarycentralnervoussystemlymphoma，PCNSL）是一种罕见的结外恶性淋巴瘤，主要亚型为弥漫性大B细胞淋巴瘤（diffuse largeB-cell lymphoma，DLBCL），占比超过 90% ，其5年生存率为 30%～40%[1] 。（剩余8689字）