Region-Based Dense Adversarial Generation for Medical Image Segmentation

Abstract

AbstractDeep neural networks (DNNs) have achieved great success in medical image segmentation. However, the DNNs are generally deceived by the adversarial examples, making robustness a key factor of DNNs when applied in the field of medical research. In this paper, in order to evaluate the robustness of medical image segmentation networks, we propose a novel Region-based Dense Adversary Generation (RDAG) method to generate adversarial examples. Specifically, our method attacks the DNNs on both pixel-level and region-of-interesting (ROI) level. The pixel-level attack makes DNNs mistakenly segment each individual pixel. Meanwhile, the ROI-level attack will generate perturbation based on region information. We evaluate our proposed method for medical image segmentation on DRIVE and CELL datasets. The experimental results show that our proposed method achieves effective attack results on both datasets for medical image segmentation when compared with several state-of-the-art methods.KeywordsAdversarial examplesMedical image segmentationRegion-based dense adversary generation