Towards Robust General Medical Image Segmentation

The reliability of Deep Learning systems depends on their accuracy but also robustness against adversarial perturbations to the input data. Several attacks and defenses have been proposed improve performance Neural Networks under presence noise in natural image domain. However, computer-aided diagnosis for volumetric data has only explored specific tasks with limited attacks. We propose a new framework assess general medical segmentation systems. Our contributions are two-fold: (i) we benchmark evaluate context Medical Segmentation Decathlon (MSD) by extending recent AutoAttack classification domain segmentation, (ii) present novel lattice architecture RObust Generic (ROG). results show that ROG is capable generalizing across different MSD largely surpasses state-of-the-art sophisticated