Overall Survival Prediction in Gliomas Using Region-Specific Radiomic Features

In this paper, we explored predictive performance of region-specific radiomic models for overall survival classification task in BraTS 2019 dataset. We independently trained three radiomic models: single-region model which included radiomic features from whole tumor (WT) region only, 3-subregions model which included radiomic features from non-enhancing tumor (NET), enhancing tumor (ET), and edema (ED) subregions, and 6-subregions model which included features from the left and right cerebral cortex, the left and right cerebral white matter, and the left and right lateral ventricle subregions. A 3-subregions radiomics model relied on a physiology-based subdivision of WT for each subject. A 6-subregions radiomics model relied on an anatomy-based segmentation of tumor-affected regions for each subject which is obtained by a diffeomorphic registration with the Harvard-Oxford subcortical atlas. For each radiomics model, a subset of most predictive features was selected by ElasticNetCV and used to train a Random Forest classifier. Our results showed that a 6-subregions radiomics model outperformed the 3-subregions and WT radiomic models on the BraTS 2019 training and validation datasets. A 6-subregions radiomics model achieved a classification accuracy of 47.1% on the training dataset and a classification accuracy of 55.2% on the validation dataset. Among the single subregion models, Edema radiomics model and Left Lateral Ventricle radiomics model yielded the highest classification accuracy on the training and validation datasets.