基于多模态磁共振成像栖息地影像组学预测低级别胶质瘤患者预后

Application of multimodal magnetic resonance imaging-based habitat radiomics in predicting the prognosis of patients with low-grade glioma

摘要:

目的 利用K-means聚类算法从磁共振成像（MRI）序列中识别异质性的功能亚区，并建立栖息地风险评分（HRS）模型预测低级别胶质瘤（LGG）患者的预后。方法 收集143例LGG患者的相关临床及影像学资料，利用无监督K-meams聚类算法对胶质瘤栖息地进行功能亚区聚类，分别提取不同功能亚区组学特征。进而构建不同亚区HRS，分析HRS与总生存时间（OS）的相关性，并对不同功能HRS进行外部验证。多因素Cox回归分析建立临床、栖息地联合临床模型，采用时间依赖的受试者操作特征（ROC）曲线，计算曲线下面积（AUC）评估不同模型对LGG患者预后预测性能。结果 基于K-meams聚类算法确定最佳分区为3个亚区，中位生存期K-M生存曲线显示，训练组基于Habitat 2亚区（高灌注高细胞致密区）构建的HRS2与OS相关（P=0.001）。多因素Cox回归分析显示年龄（HR=1.033）、WHO分级（HR=1.290）、HRS2（HR=2.498）是预测LGG预后的影响因素。基于以上结果建立栖息地联合临床预测模型，并对模型进行外部验证。训练组队列临床、栖息地联合临床模型预测LGG患者OS的AUC分别为0.711、0.855，而验证组队列AUC分别为0.709、0.857。结论 生境技术可以通过分割肿瘤不同亚区量化肿瘤异质性，基于高危亚区构建的HRS是LGG患者预后的影响因素，栖息地联合临床模型在预后评估方面优于临床模型。

Abstract:

Objective To identify the functional subregions characterizing tumor heterogeneity from magnetic resonance imaging （MRI） sequence using the K-means clustering algorithm， and to construct a habitat risk score （HRS） model for predicting the prognosis of patients with low-grade glioma （LGG）.Methods Clinical and imaging data were collected from 143 patients with LGG. The unsupervised K-means clustering algorithm was used to cluster the functional subregions of glioma habitats， and the radiomic features of different subregions were extracted. HRS was established for different subregions， and its correlation with overall survival （OS） was analyzed. External validation was performed for HRS. The multivariate Cox regression analysis was used to establish a clinical model and a habitat-clinical model， and the time-dependent receiver operating characteristic （ROC） curve was used to assess the performance of different models in predicting the prognosis of LGG patients.Results The K-means clustering algorithm identified the optimal partition of 3 subregions， and the Kaplan-Meier survival curve for median survival time showed that HRS2 constructed based on Habitat 2 subregion （the area with high perfusion and cellular density） was significantly associated with OS （P = 0.001）. The multivariate Cox regression analysis showed that age （hazard ratio ［HR］ = 1.033）， WHO grade （HR = 1.290）， and HRS2 （HR= 2.498） were influencing factors for the prognosis of LGG. A habitat-clinical model was established based on the above results， and external validation was performed for this model. For the training cohort， the clinical model and the habitat-clinical model had an area under the ROC curve （AUC） of 0.711 and 0.855， respectively， in predicting the OS of LGG patients， while in the validation cohort， the two models had an AUC of 0.709 and 0.857， respectively.Conclusions Habitat technology can quantify tumor heterogeneity by segmenting different tumor subregions. HRS developed based on high-risk subregions is an influencing factor for the prognosis of LGG， and the habitat-clinical model has a better effect than the clinical model in prognostic assessment.