基于双参数磁共振影像组学模型预测老年前列腺癌患者根治术后生化复发

Biparametric magnetic resonance imaging radiomics for predicting biochemical recurrence in elderly prostate cancer patients after radical prostatectomy

目的探讨基于双参数磁共振(bpMRI)影像组学模型对老年(≥60岁)前列腺癌患者根治术(RP)后生化复发(BCR)的预测价值。方法回顾性分析2017年8月至2021年12月北京医院收治的175例患者的资料,根据病理结果分别在术前bpMRI的T2、磁共振扩散加权成像(DWI)和表观弥散系数(ADC)序列上进行图像分割。使用Pyradiomics提取影像组学特征,采用Cox回归、Spearman相关系数和LASSO回归对特征进行降维,构建影像组学标签。采用多因素Cox回归分析构建临床模型和影像-临床联合模型,使用一致性指数(C-index)评价模型预测BCR的效能。结果175例患者按照7∶3随机分为训练集(122例)和测试集(53例),分别有24例(19.7%、24/122)和11例(20.8%、11/53)患者发生BCR。在3个序列中提取5775个影像组学特征,最终筛选5个特征构建影像组学标签。影像组学模型在训练集和测试集的C-index分别为0.764(95%CI:0.655~0.872)和0.769(95%CI:0.632~0.906)。多因素Cox回归分析结果显示,血清前列腺特异性抗原(PSA)(HR=1.032,95%CI:1.010~1.054)、术后病理国际泌尿外科病理学学会(ISUP)分级(HR=1.682,95%CI:1.039~2.722)和切缘阳性(HR=2.513,95%CI:1.094~5.774)是BCR的独立预测因子。临床模型在训练集和测试集的C-index分别为0.751(95%CI:0.655~0.846)和0.753(95%CI:0.630~0.877)。联合临床因素和影像组学标签建模后,影像-临床联合模型的C-index最高,分别为0.782(95%CI:0.679~0.874)和0.801(95%CI:0.677~0.915)。结论基于bpMRI构建的影像组学模型可以预测老年前列腺癌患者RP术后BCR的发生,联合临床因素和影像组学标签建模可以提高预测效能。

Objective To investigate the predictive value of a radiomics model based on biparametric magnetic resonance imaging(bpMRI)for biochemical recurrence(BCR)after radical prostatectomy(RP)in elderly prostate cancer patients(≥60 years old).MethodsA retrospective analysis was conducted on data from 175 patients treated at Beijing Hospital from August 2017 to December 2021.Based on pathological results,image segmentation was performed on preoperative bpMRI T2,diffusion weighted imaging(DWI),and apparent diffusion coefficient(ADC)sequences.Pyradiomics was utilized to extract radiomic features,and Cox regression,Spearman correlation coefficient,and LASSO regression were employed for feature dimensionality reduction,leading to the construction of radiomic labels.Clinical models and image-clinical combined models were developed using multifactorial Cox regression analysis,and the performance of these models in predicting BCR was evaluated using the concordance index(C-index).ResultsThe 175 patients were randomly divided into a training set(122 cases)and a test set(53 cases)at a ratio of 7∶3,with 24 cases(19.7%,24/122)and 11 cases(20.8%,11/53)experiencing BCR,respectively.A total of 5775 radiomic features were extracted from the three sequences,and after dimensionality reduction,5 features were selected to construct the radiomic labels.The radiomics model exhibited C-index values of 0.764(95%CI:0.655-0.872)and 0.769(95%CI:0.632-0.906)in the training and test sets,respectively.Multifactorial Cox regression analysis revealed serum prostate-specific antigen(PSA)(HR=1.032,95%CI:1.010-1.054),postoperative pathology International Society of Urological Pathology(ISUP)grade grouping(HR=1.682,95%CI:1.039-2.722),and positive surgical margins(HR=2.513,95%CI:1.094-5.774)as independent predictors of BCR.The clinical model exhibited C-index values of 0.751(95%CI:0.655-0.846)and 0.753(95%CI:0.630-0.877)in the training and test sets,respectively.Following combined modeling of clinical factors and radiomic labels,the image-clinical combined model demonstrated the highest C-index values,namely 0.782(95%CI:0.679-0.874)and 0.801(95%CI:0.677-0.915)in the training and test sets,respectively.ConclusionsThe radiomics model based on bpMRI can predict the occurrence of BCR after RP in elderly prostate cancer patients.Combined modeling of clinical factors and radiomic labels can enhance predictive efficiency.