Challenging the Robustness of Image Registration Similarity Metrics with Adversarial Attacks

Rexeisen, Robin; Jiang, Xiaoyi

Abstract

Multi-modal image registration is a crucial task in various medical applications. A typical technique here is iterative optimization, whose success depends on the reliability of the used similarity metric. In this work, we systematically challenge the robustness of two such popular metrics, Mutual Information and Cross-Cumulative Residual Entropy, by employing adversarial techniques from the deep learning field. Our experiments show resistance to small perturbations, while indicating a higher vulnerability as the perturbations increase. Furthermore, our results indicate that certain structural patterns emerge during this process. Additionally, we examine the functional landscape of both metrics. Consequently, this work emphasizes the robustness of these metrics while also offering a starting point for more insights into the underlying patterns that contribute to their failure.