This study compared the predictive power and robustness of texture, topological, and convolutional neural network (CNN) based image features for measuring tumors in MRI. These features were used to predict 1p/19q codeletion in the MICCAI BRATS 2017 challenge dataset. Topological data analysis (TDA) based on persistent homology had predictive performance as good as or better than texture-based features and was also less susceptible to image-based perturbations. Features from a pre-trained convolutional neural network had similar predictive performances and robustness as TDA, but also performed better using an alternative classification algorithm, k-top scoring pairs. Feature robustness can be used as a filtering technique without greatly impacting model performance and can also be used to evaluate model stability.
Keywords: 1p/19q codeletion; glioma; image perturbation; multiparametric MRI; persistent homology; radiomic features.