Radiomics-guided prediction of overall survival (OS) in brain gliomas is seen as a significant problem in Neuro-oncology. The ultimate goal is to develop a robust MRI-based approach (i.e., a radiomics model) that can accurately classify a novel subject as a short-term survivor, a medium-term survivor, or a long-term survivor. The BraTS 2020 challenge provides radiological imaging and clinical data (178 subjects) to develop and validate radiomics-based methods for OS classification in brain gliomas. In this study, we empirically evaluated the efficacy of four multiregional radiomic models, for OS classification, and quantified the robustness of predictions to variations in automatic segmentation of brain tumor volume. More specifically, we evaluated four radiomic models, namely, the Whole Tumor (WT) radiomics model, the 3-subregions radiomics model, the 6-subregions radiomics model, and the 21-subregions radiomics model. The 3-subregions radiomics model is based on a physiological segmentation of whole tumor volume (WT) into three non-overlapping subregions. The 6-subregions and 21-subregions radiomic models are based on an anatomical segmentation of the brain tumor into 6 and 21 anatomical regions, respectively. Moreover, we employed six segmentation schemes – five CNNs and one STAPLE-fusion method – to quantify the robustness of radiomic models. Our experiments revealed that the 3-subregions radiomics model had the best predictive performance (mean AUC = 0.73) but poor robustness (RSD = 1.99) and the 6-subregions and 21-subregions radiomics models were more robust (RSD 1.39) with lower predictive performance (mean AUC 0.71). The poor robustness of the 3-subregions radiomics model was associated with highly variable and inferior segmentation of tumor core and active tumor subregions as quantified by the Hausdorff distance metric (4.4−6.5mm) across six segmentation schemes. Failure analysis revealed that the WT radiomics model, the 6-subregions radiomics model, and the 21-subregions radiomics model failed for the same subjects which is attributed to the common requirement of accurate segmentation of the WT volume. Moreover, short-term survivors were largely misclassified by the radiomic models and had large segmentation errors (average Hausdorff distance of 7.09mm). Lastly, we concluded that while STAPLE-fusion can reduce segmentation errors, it is not a solution to learning accurate and robust radiomic models.
