研究通过分析放疗期间不同时间点的非增强FLAIR高信号区域的放射组学特征，发现这些特征能够预测胶质母细胞瘤患者的无进展生存期，并进行风险分层。利用统计等效签名算法和混合效应Cox回归，从1743个特征中构建预测模型。结果表明，结合多个时间点MRI的放射组学签名在验证组中表现出良好的预测性能，对于患者风险分层具有重要意义。
摘要：
Purpose: To evaluate whether non-enhancing FLAIR hyperintense radiomic features generated from MRIs performed at several timepoints during radiation therapy (RT) can predict progression free survival (PFS) and/or risk stratification of patients with newly diagnosed glioblastoma (GBM). Materials and Methods: 35 patients underwent 4 MRIs at: RT planning (D0); on the 10th (D10) and 20th (D20) fractions of RT; and at one-month post-RT follow-up (D70). Endpoints were based on clinical and imaging RANO criteria at 6 months after RT. All neoplastic-related FLAIR hyperintense regions on all MRIs were contoured using a threshold-based method, with T1 contrast enhancing and necrotic regions excluded. 1743 features were extracted using pyradiomics [1], then scaled and centered. Patient datasets were randomly split into training (n=24) and validation (n=11) datasets. Statistically Equivalent Signatures (SES) algorithm equipped with generalized linear mixed modelling (GLMM) and mixed effects Cox regression were used for feature selection [2]. Radiomic signatures for each patient was constructed through linear combination of features weighted by their model coefficients. Radiomics score was calculated as median of signature values for each patient. The median radiomics score was used as a threshold for stratifying patients into high- vs low- risk groups. This threshold was tested in the validation group, and Kaplan-Meier and log-rank tests were used to test the difference between the resulting high- and low- risk groups. The concordance index was used to assess the performance of the signature in the validation dataset. Results: All models using one (D0) or series of MRI images (D0-D10, D0-D20, and D0-D70) were significantly predictive of PFS in the training dataset (p<0.001). Each additional MRI improved model performance in the validation dataset, with the D0-D70 model achieving the best performance for patient risk stratification with a log-rank score of 6.21 (p=0.01) and hazard ratio of 9.966 (95% CI: 1.127, 88.14, p=0.04). This model also had the best concordance index at 0.75 (95% CI: 0.691, 0.809) for the validation dataset. Conclusions: The radiomics signature constructed from non-enhancing FLAIR hyperintense regions of tumor from D0-D70 MRIs is an independent predictor of PFS in patients with GBM. Novel radiomics biomarkers acquired from non-invasive longitudinal MRI images can have a significant impact in care of GBM patients.
展开
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
