Adversarial perturbation on MRI modalities in brain tumor segmentation

Abstract

Convolutional neural networks (CNNs) have been widely used by biomedical image segmentation applications. U-net, as a semantic segmentation method, has become a mainstream approach to brain tumor segmentation. However, the intrinsic vulnerability of CNNs also brings potential risks to all CNN-based applications, including semantic segmentation applications. In this paper, we create a universal adversarial perturbation and apply it on every modality in order to investigate how the adversarial perturbation affects each Magnetic Resonance Imaging(MRI) modality and the MRI images overall. We evaluate the performance when all four modalities are attacked and when one modality is attacked. The results show the following: 1) The adversarial perturbation affects the accuracy performance greatly, regardless of the size of the perturbation; 2) When only one modality is attacked, the network structure and the other three modalities provide some resistance to the adversarial perturbation; and 3) There are performance differences in different modalities, which are strongly related to the intensity distribution. T2 is least affected by the adversarial perturbation, while T1 and T1ce are more affected by the adversarial perturbation.