Robustifying Deep Networks for Medical Image Segmentation.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Deep Learning for Medical Image Analysis

Background:

• Deep learning models, particularly convolutional neural networks (CNNs), are increasingly used for medical image segmentation.
• The robustness of these models against adversarial perturbations, which are subtle, imperceptible image modifications, is a critical concern for clinical reliability.

Purpose of the Study:

• To investigate the vulnerability of a common CNN (UNet) for brain tumor segmentation to adversarial perturbations.
• To evaluate the effectiveness of adversarial attack strategies and defense mechanisms, including distillation and adversarial training.

Main Methods:

• Implemented two UNet models for segmenting four MRI series (T1, T1-contrast, T2, T2-FLAIR) of low- and high-grade gliomas.
• Developed adversarial attacks using Fast Gradient Sign Method (FGSM), iterative FGSM (i-FGSM), and targeted iterative FGSM (ti-FGSM).
• Assessed defense strategies: distillation and adversarial training via data augmentation; robustness measured by Dice coefficients.

Main Results:

• Adversarial attacks effectively reduced segmentation quality, with Dice coefficients decreasing by up to 65%.
• Distillation demonstrated superior performance over adversarial training in mitigating attack effects.
• All defense methods resulted in lower performance compared to unperturbed images, indicating limitations in current defenses.

Conclusions:

• Segmentation networks are susceptible to targeted adversarial attacks with visually minor modifications, impacting diagnostic accuracy.
• Quantifying the impact of adversarial inputs is crucial for the safe and reliable application of deep learning in medical imaging.
• While distillation offers some protection, current defense strategies do not fully restore performance against sophisticated adversarial attacks.