基于T2加权成像的影像组学列线图模型预测前列腺癌盆腔淋巴结转移

Prediction of pelvic lymph node metastasis in prostate cancer using radiomics based on T2-weighted imaging

目的:盆腔淋巴结转移是影响前列腺癌分期及预后的重要因素,有创的扩大盆腔淋巴结清扫(extended pelvic lymph node dissection,e PLND)被认为是临床上确诊淋巴结转移的最有效方法,术前准确预测盆腔淋巴结转移可减少不必要的ePLND。本研究旨在探讨基于T加权成像(T-weighted imaging,TWI)的影像组学列线图预测前列腺癌盆腔淋巴结转移的临床价值。方法:回顾性收集北京大学第一医院2017年1月至2021年6月的71例行ePLND的前列腺癌患者的磁共振(magnetic resonance,MR)资料。所有患者按照多参数磁共振成像(multiparametric magnetic resonance imaging,mpMRI)扫描时间分为训练集(2017年1月至2020年12月,n=56,包含186个淋巴结)和测试集(2021年1月至2021年6月,n=15,包含45个淋巴结)。由2名放射科医师对清扫淋巴结与MRI图像进行匹配并勾画感兴趣区(region of interest,ROI)。基于勾画的ROI建立模型1(单独TWI的图像特征)、模型2(随机森林筛选的影像组学特征)及模型3(组合的图像特征和组学特征)3个转移淋巴结的预测模型并建立列线图。从病历记录中收集所有患者的临床病理特征,包括年龄、Gleason评分、前列腺特异性抗原(prostate specific antigen,PSA)水平以及临床和病理T分期。测量手工勾画的淋巴结(包括转移淋巴结和非转移淋巴结)长径、短径及体积。采用受试者工作特征(receiver operating characteristic,ROC)曲线评估3个模型的诊断效能,决策曲线分析(decision curve analysis,DCA)评估模型的临床收益。结果:训练集和测试集中患者的年龄、Gleason评分、PSA水平、临床和病理T分期的差异均无统计学意义(均P>0.05)。而转移淋巴结和非转移淋巴结的体积、短径和长径的差异在训练集和测试集中均有统计学意义(均P<0.05)。在多因素回归分析中,淋巴结的短径和边缘状况被纳入构建模型1。18个组学特征经筛选后用于构建模型2。淋巴结信号分布和Rad score为模型3预测盆腔淋巴结转移的显著危险因素。基于模型3建立的列线图C-指数达0.964,校准曲线显示该模型具有较高的校准度。在测试集中,模型1,2,3的ROC曲线下面积(area under the curve,AUC)分别为0.78,0.93,0.96,模型2和模型3的诊断效能均显著高于模型1(模型1 vs模型2,P=0.019;模型1 vs模型3,P=0.020),而模型2与模型3 AUC值的差异无统计学意义(P=0.649)。3个模型的DCA表明:在不同的阈值概率范围内,3个模型都获得了比全(认为全部淋巴结都为转移淋巴结)或无(认为无转移淋巴结)方案更高的收益,其中模型3具有最高的临床获益。结论:基于TWI的影像组学列线图对前列腺癌患者盆腔淋巴结转移具有良好的预测效能,可作为预测淋巴结转移的影像学标志物,指导进行淋巴结清扫及制订随访策略。

Objective: Pelvic lymph node metastasis(PLNM) is an important factor that affects the stage and prognosis of prostate cancer. Invasive extended pelvic lymph node dissection(ePLND) is the most effective method for clinically diagnosing PLNM. Accurate preoperative prediction of PLNM can reduce unnecessary ePLND. This study aims to investigate the clinical value of radiomics nomogram in predicting PLNM of prostate cancer based on T-weighted imaging(TWI).Methods: Magnetic resonance(MR) data of 71 patients with prostate cancer who underwent ePLND from January 2017 to June 2021 in Peking University First Hospital were collected retrospectively. All patients were assigned into a training set(January 2017to December 2020, n=56, containing 186 lymph nodes) and a test set(January 2021 to June2021, n=15, containing 45 lymph nodes) according to the examination time of multiparametric magnetic resonance imaging(mpMRI). Two radiologists matched the dissected lymph nodes on MRI images, and manually annotated the region of interest(ROI). Based on the outlined ROI, 3 metastatic lymph node prediction models were established: Model 1(only image features of TWI), Model 2(radiomics features based on random forest), and Model 3(combination of the image and radiomics features). A nomogram was also established. The clinicopathologic characteristics of the patients were obtained from the medical records, including age, the Gleason score, the level of prostatespecific antigen(PSA), and clinical and pathological T stage. The preoperative radiological features of the pelvic lymph nodes(LNs) include size of LNs(the short and long diameters)and volume of LNs. Receiver operating characteristic(ROC) curve was used to evaluate the diagnostic efficacy of the 3 models and decision curve analysis(DCA) was used to evaluate the clinical benefits of the models.Results: No significant differences were found between the training set and test set regarding age, Gleason scores, PSA level, and clinical and pathological T stage(all P>0.05). The differences in volume, short diameter and long diameter between metastatic and non-metastatic LNs were statistically significant in both training set and test set(all P<0.05). In multivariate regression analysis, the short diameter and marginal status of LNs were included in Model 1. Eighteen omics features were selected to construct Model 2. The signal distribution of LNs and Rad score were the significant risk factors for predicting metastasis of pelvic LNs in Model 3. The C-index of nomogram based on Model 3 reached0.964, and the calibration curve showed that the model had high calibration degree. In the test set, the area under the curves of Model 1, 2, and 3 were 0.78, 0.93, and 0.96respectively, Model 2 and Model 3 showed significantly higher diagnostic efficiency than Model 1(Model 1 vs Model 2, P=0.019;Model 1 vs Model 3, P=0.020). There was no significant difference in the area under the curve between Model 2 and Model 3(P=0.649).The DCA results of the 3 models showed that all models obtained higher net benefits than the PLNM-all or PLNM-none protocol in different ranges of threshold probabilities and Model 3 had the highest clinical benefit.Conclusion: The radiomics nomogram based on TWI shows a good predictive efficacy for preoperative PLNM in patients with prostate cancer, which could be served as an imaging biomarker to optimize decision-making and adjust adjuvant treatments.