Adversarial attacks and adversarial training for burn image segmentation based on deep learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Deep learning models are susceptible to adversarial attacks, particularly in medical image segmentation.
• Medical images often contain noise due to imaging conditions, impacting model performance.
• Existing adversarial training methods may not fully address physically meaningful disturbances in burn images.

Purpose of the Study:

• To propose an adversarial training approach for burn image segmentation that simulates natural phenomena-inspired attacks.
• To enhance the robustness and accuracy of deep learning models in segmenting burn images.
• To reduce the computational cost associated with adversarial training.

Main Methods:

• Simulated adversarial attacks inspired by natural phenomena.
• Developed a specialized adversarial training approach for burn image segmentation.
• Tested the method on a dedicated burn image dataset.
• Conducted ablation experiments to validate individual loss components.

Main Results:

• Achieved a segmentation accuracy of 82.19% for adversarial samples, an increase from the initial 54%.
• Demonstrated a 1.97% improvement over conventional adversarial training methods.
• Significantly reduced training time compared to standard approaches.
• Validated the effectiveness of proposed losses and compared performance across different adversarial samples.

Conclusions:

• The proposed adversarial training method effectively improves deep learning model performance for burn image segmentation.
• The approach enhances robustness against naturalistic image disturbances while being computationally efficient.
• This technique offers a promising direction for reliable medical image analysis in challenging conditions.