Toward Robust and Generalizable Radio(gen)omics Predictive Models for Brain Tumor Characterization

In the context of brain tumor characterization, we focused on 2 key questions that, to the best of our knowledge, have not been explored so far: 1) stability of radiomics features to variability in multiregional segmentation masks obtained with fully automatic deep segmentation methods and 2) subsequent impact on predictive performance on downstream prediction tasks. The hypothesis is that highly stable and discriminatory radiomics features lead to generalizable radiogenomics models in brain tumor characterization.

We used the publicly available Brain Tumor Segmentation 2020 data set for tumor segmentation and isocitrate dehydrogenase (IDH) prediction. For segmentation, the training cohort included 369 subjects with preoperative multiparametric 3D MRI (T1, T1-Gadolinium enhanced [T1-Gd], T2, and FLAIR) and manual annotations of tumor subregions (whole tumor, WT; tumor core, TC; enhancing core, EC), while the validation cohort comprised 125 subjects with imaging data only. For IDH prediction, the discovery data set consisted of 148 subjects (57 IDH-mutant, 91 IDH-wild-type) and the testing data set included 70 subjects (32 IDH-mutant, 38 IDH-wild-type). Seven state-of-the-art convolutional neural networks were used for fully automatic multiregional tumor segmentation. Radiomics feature stability across segmentation models was assessed using the overall concordance correlation coefficient (OCCC), and discriminatory features were selected with recursive feature elimination with support vector machines. Predictive performance was evaluated using area under the receiver operating characteristic curve (AUROC), and model stability was quantified by the relative SD (RSD) of AUROC.

Our study found that highly stable radiomics features were predominantly texture-based (79.1%), mainly extracted from the WT region (96.1%), and largely derived from T1-Gd (35.9%) and T1 (28.0%) sequences. Mean feature stability (OCCC) was highest for WT (0.87 ± 0.12), followed by TC (0.76 ± 0.13), EC (0.72 ± 0.13), and shape features (0.72 ± 0.11), with shape and EC features showing the lowest stability. Stability filtering reduced nonphysiologic variability, as reflected by a lower RSD (2.28% versus 0.64%), and significantly improved predictive performance across 8 segmentation schemes (AUROC: 0.81 ± 0.02 versus 0.94 ± 0.006).

Robust and generalizable radiogenomics models can be learned with highly stable and discriminatory radiomics features.