Elevating adversarial robustness by contrastive multitasking defence in medical image segmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep Learning (DL) models are crucial for Medical Image Segmentation (MIS).
• Adversarial attacks degrade DL-based MIS model performance and robustness.
• Existing defenses are less effective in the medical domain.

Purpose of the Study:

• To propose a novel defense strategy, CEASE (Contrastive Multitasking Defense), to enhance adversarial resilience in DL-based MIS models.
• To investigate the efficacy of contrastive and multitask learning for improving robustness against adversarial attacks in medical imaging.

Main Methods:

• CEASE integrates contrastive learning to ensure similar feature learning for clean, adversarial, and augmented samples.
• A multitask learning approach is employed to generate generic feature representations and improve robustness.
• A fusion-based defense consolidates contrastive and multitask learning for optimal adversarial resilience.

Main Results:

• CEASE significantly enhances adversarial resilience in state-of-the-art MIS models.
• The proposed defense mitigates adversarial attacks, achieving up to 0% attack success rate.
• CEASE demonstrates modest performance gains while improving robustness.

Conclusions:

• CEASE effectively bridges the gap in adversarial defense for medical image segmentation.
• The integration of contrastive and multitask learning offers a promising direction for robust DL models in healthcare.
• CEASE provides a robust solution for defending medical image segmentation models against adversarial attacks.