Jingjing Shi1, Shaowei Yang2, Jian Wang1, Sui Huang2, Yihao Yao1, Shun Zhang1, Wenzhen Zhu3, Jianbo Shao4. 1. Department of Radiology, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China. 2. Department of Radiology, Children's Hospital, Wuhan, China. 3. Department of Radiology, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China. Electronic address: zhuwenzhen8612@163.com. 4. Department of Radiology, Children's Hospital, Wuhan, China. Electronic address: shaojb2002@sina.com.
Abstract
PURPOSE: On MR imaging, peritumoral T2 hyperintensity surrounding glioblastoma is known to contain tumor cell infiltrates, thus contributing to poor prognosis. This study aimed to determine the incremental prognostic value of radiomics on peritumoral T2 hyperintensity in pretreatment glioblastoma. METHODS: One hundred fourteen pathologically confirmed glioblastoma patients were retrospectively selected from March 2008 to May 2018 (our institution, n = 61; the Cancer Imaging Archive, n = 53). All patients were randomly divided into either training (n = 80) or test set (n = 34). Manually segmented peritumoral T2 hyperintensity yielded 106 radiomic features per patient. A random forest variable selection was used to select the most relevant radiomic features. Four Cox proportional hazards models were fitted with clinical features, clinical features with tumor/peritumoral volumes, radiomics, and all of them combined. Kaplan-Meier survival curves of the models were plotted with log-rank tests. All models were validated on a test set using prediction error curves over survival times. RESULTS: A random forest variable selection yielded five relevant features among the 106 radiomic features (two shape, two gray-level and one first order features). These radiomic features increased survival prediction accuracy when they were added onto clinical and tumor/peritumoral volumetric features (combined model, P = 0.011). On test set, the combined model showed lower mean survival prediction error rate (0.14) than clinical (0.191) or radiomic (0.178) model. CONCLUSIONS: The clinical model with radiomic features demonstrated improved survival predictive performance than the model without radiomic features, thus suggesting incremental prognostic value of peritumoral radiomics as MR imaging biomarker in pretreatment glioblastoma.
PURPOSE: On MR imaging, peritumoral T2 hyperintensity surrounding glioblastoma is known to contain tumor cell infiltrates, thus contributing to poor prognosis. This study aimed to determine the incremental prognostic value of radiomics on peritumoral T2 hyperintensity in pretreatment glioblastoma. METHODS: One hundred fourteen pathologically confirmed glioblastomapatients were retrospectively selected from March 2008 to May 2018 (our institution, n = 61; the Cancer Imaging Archive, n = 53). All patients were randomly divided into either training (n = 80) or test set (n = 34). Manually segmented peritumoral T2 hyperintensity yielded 106 radiomic features per patient. A random forest variable selection was used to select the most relevant radiomic features. Four Cox proportional hazards models were fitted with clinical features, clinical features with tumor/peritumoral volumes, radiomics, and all of them combined. Kaplan-Meier survival curves of the models were plotted with log-rank tests. All models were validated on a test set using prediction error curves over survival times. RESULTS: A random forest variable selection yielded five relevant features among the 106 radiomic features (two shape, two gray-level and one first order features). These radiomic features increased survival prediction accuracy when they were added onto clinical and tumor/peritumoral volumetric features (combined model, P = 0.011). On test set, the combined model showed lower mean survival prediction error rate (0.14) than clinical (0.191) or radiomic (0.178) model. CONCLUSIONS: The clinical model with radiomic features demonstrated improved survival predictive performance than the model without radiomic features, thus suggesting incremental prognostic value of peritumoral radiomics as MR imaging biomarker in pretreatment glioblastoma.