AdverIN: Monotonic adversarial intensity attack for domain generalization in medical image segmentation

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Domain generalization (DG) aims to improve deep learning model performance on unseen data domains.
• Current DG methods focus on learning domain-invariant features for robustness.
• Handling variations across different data domains remains a significant challenge in AI.

Purpose of the Study:

• To propose a novel domain generalization technique called Adversarial Intensity Attack (AdverIN).
• To enhance the generalizability of deep learning models for medical image segmentation tasks.
• To improve model robustness against domain shifts by synthesizing diverse image intensities.

Main Methods:

• Adversarial Intensity Attack (AdverIN) employs an adversarial training strategy.
• The method synthesizes a spectrum of intensity variations to augment data diversity.
• Essential contextual information within images is preserved during synthesis.

Main Results:

• AdverIN significantly improved the generalizability of segmentation models in experiments.
• State-of-the-art performance was achieved on 2D retinal optic disc/cup and 3D prostate MRI segmentation.
• The proposed technique demonstrated effectiveness across diverse multi-domain tasks.

Conclusions:

• Adversarial Intensity Attack (AdverIN) is an effective domain generalization technique.
• The method enhances model robustness and performance on previously unseen data domains.
• AdverIN shows promise for improving medical image analysis and other DG applications.