Adversarial Robustness Evaluation of Deep Learning Segmentation Models and Loss Functions in Prostate MRI

Kosmas Κ. Apostolidis, Dimitrios I. Zaridis, Nikolaos S. Tachos, Vasileios C. Pezoulas, Kostas Marias, Manolis Tsiknakis, Dimitrios I. Fotiadis

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Adversarial robustness, Adversarial Attacks, Medical Image Segmentation, Prostate MRI, Deep Learning

TL;DR: This paper examines the adversarial robustness of five deep learning segmentation model architectures and six loss functions in a prostate MRI dataset under an computationally efficient white-box adversarial attack.

Abstract: Deep learning has significantly progressed the field of medical image segmentation; yet, its susceptibility to adversarial attacks affects clinical and end-user confidence in the automated solutions derived from the AI systems. This study examines the adversarial robustness of five deep learning-based segmentation models and six loss functions against the Fast Gradient Sign Method attack across multiple attack strengths on a prostate MRI dataset. The experimental findings indicate that Recurrent U-Net had the highest adversarial robustness. Specifically, it surpassed all the other assessed models in two out of three evaluation metrics, achieving a mean of 0.52 Dice Coefficient and a 95th percentile Hausdorff Distance of 11.12 across the folds. Additionally, in the hold-out dataset, it attained a mean of 0.54 Dice Coefficient, a 95th percentile Hausdorff Distance of 9.45 and an Average Surface Distance of 0.93. Likewise, the loss functions derived from the Tversky index had the highest adversarial robustness. Precisely, Tversky loss surpassed all the other assessed loss functions in two out of three metrics across the folds, with a mean of 0.57 Dice Coefficient and 13.07 for the 95th percentile Hausdorff Distance and in all three evaluation metrics in the hold-out dataset, with a mean of 0.52 Dice Coefficient, a 95th percentile Hausdorff Distance of 10.05 and an Average Surface Distance of 0.74 when combined with the Binary Cross-Entropy loss function. From a clinical perspective, the findings of this work can guide the development of more adversarially resilient AI segmentation systems.

Track: 2. Bioinformatics

Registration Id: 3GNVCGQWTYV

Submission Number: 49