Purpose: Accurate survival prediction may assist clinic-decision-making for personalized treatment for glioma. We aimed to develop a novel framework, consisting of a two-phase feature selection and a fused regression forest model, to achieve robust survival prediction of patients with glioma while reducing the computational cost.
Methods: BraTS19 211 high-grade patients with glioma were included. The patients were stratified into three subgroups according to the reported survival intervals: long-term (>15 months), mid-term (10-15 months), and short-term(<10 months). Each patient had four preoperative multimodal MR scans (T1, T1ce, T2 and Flair). Each tumor was further divided into three subregions: non-enhancing tumor (NET), enhancing tumor (ET) and peritumoral edema (ED). Instead of extraction of all features from all four MRIs, we first extracted a few robust features (18 first-order and 22 geometric features) from each subregion on the MR sequence that is most relevant (or optimized out of 64 combinations) for each subregion using a clustering evaluation method. Additional 1632 high-order texture features were then extracted from their corresponding MRI sequences. Recursive feature elimination algorithm was performed to generated 3 different feature sets for three subgroups, respectively. Each feature set was used to train an independent random forest (RF) model for each patient subgroup. The final prediction model was formed by fused the RF models.
Results: The corresponding relationship that is most conducive to feature extraction between each tumor subregion and MRI sequence were determined. The optimized imaging feature sets were determined for patient survival subgroups, respectively. The proposed framework achieved an Area Under the ROC Curve (AUC) of 0.722 with least mean square error using a 10-Fold cross-validation.
Conclusion: We developed a novel framework for tumor subregional based treatment response prediction on multiparametric MRIs. The proposed framework effectively identified an optimized feature set and achieved superior survival prediction for patients with glioma.
Funding Support, Disclosures, and Conflict of Interest: The Natural Science Basic Research Program of Shaanxi Province of China [2019JM-365]; The Scientific Research Program Funded by Shaanxi Provincial Education Department of China[17JK0701]; The Postgraduate Innovation Fund of Xi'an University of Posts and Telecommunications[CXJJLZ202012]
MRI, Quantitative Imaging, Feature Selection
TH- Response Assessment: Radiomics/texture/feature-based response assessment