基于多模态MRI栖息地影像组学预测较低级别胶质瘤患者预后

Predicting survival in lower grade gliomas patients using multimodal MRI-based habitat radiomics

摘要:

目的：利用K-means聚类算法从MRI中识别具有肿瘤异质性的功能亚区，并基于亚区建立栖息地风险评分（habitat risk score，HRS）模型预测LGGs患者预后。 方法：收集143例LGGs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，分别提取不同亚区组学特征。进而构建不同亚区栖息地风险评分（habitat risk score，HRS），分析栖息地风险评分与总生存期（overall survival，OS）的相关性，并对不同栖息地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）曲线评估不同模型对LGGs患者预后的预测性能。 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生存曲线发现训练组基于Habitat 2亚区（高灌注高细胞致密区）构建的HRS2与OS显著相关（p=0.001）。多因素Cox发现年龄(HR=1.033)、WHO分级(HR=1.29)、HRS2(HR=22.498) 是预测LGGs预后的独立危险因素。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LGGs患者OS的AUC分别为0.711、0.855。在验证组队列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LGGs患者预后的独立危险因素，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 functional subregions characterizing tumor heterogeneity from MRI using K-means clustering algorithm, and to develop habitat models based on subregions predicting LGGs patients’ survival outcomes. Methods: A total of 143 LGGs patients from the TCGA database and First Hospital of Qinhuangdao were enrolled. Unsupervised K-means clustering algorithm was performed to cluster functional sub-regions on T1WI-enhanced and ADC images. Radiomic features were extracted from different sub-regions, then the habitat risk score (HRS) of different sub-regions was constructed, association of HRS and overall survival (OS) was analyzed in training and validation groups. Multivariate Cox regression analysis was used to establish prognostic models including clinical model and habitat-clinical model. The performance of different models in predicting the prognosis of LGGs patients was assessed using time-dependent receiver operating characteristic (ROC) curves. Results: The K-means clustering algorithm identified the optimal partition of LGGs patients into 3 sub-regions. The median survival time K-M survival curves revealed that only Habitat 2 sub-region (high perfusion and high cellular density area) was significantly associated with OS(p=0.001). Multivariate Cox analysis identified age, WHO grade, and HRS2 as independent risk factors for LGGs prognosis. Based on the three variables, Habitat-clinical model was established and verified externally. The AUC of OS predicted by the clinical model and Habitat-clinical model in the training group were 0.711 and 0.855, respectively., the AUC was 0.709 and 0.857 in the validation group. In all cohorts, the combined habitat model enhanced the diagnostic performance of the clinical model. Conclusion: Habitat technology can quantify tumor heterogeneity by segmenting different tumor sub-regions. The HRS developed based on high-risk sub-regions was an independent risk factor for LGGs prognosis, outperforming clinical models in prognostic assessment.