What Do Untargeted Adversarial Examples Reveal in Medical Image Segmentation?

Recent literature point out overconfidence problems in DNNs which is demonstrated as biased confidences in false predictions in medical image segmentation tasks regardless of the ground truth. To explore and identify the uncertain regions, we propose a post-training method with untargeted adversarial examples where the input image is iteratively perturbed in a direction that maximizes the loss of original and perturbed prediction. The perturbed predictions from these adversarial examples can be seen as unstable areas in terms of input variability; we theoretically observe that the gradient of negative class confidence in terms of input image plays a key role for perturbed outputs, and empirically show that a small adversarial perturbation can help find hidden regions in the output segmentation maps. Compared to previous methods for uncertainty estimation, our method yields competitive results for uncertain region findings on medical image datasets while only requiring one extra inference from a pre-trained model and short iteration of attack. We expect our novel findings can provide insights for future medical image segmentation problems where detection of subtle variations (e.g., lesions) are required.