Teacher-Student Instance-Level Adversarial Augmentation for Single Domain Generalized Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Single-source domain generalization (SDG) is crucial for medical image segmentation.
• Adversarial image augmentation generates challenging synthetic data but often suffers from limited diversity due to simple, image-level augmenters.
• Existing methods risk over-augmentation, hindering model generalization.

Purpose of the Study:

• To propose a novel Teacher-Student Instance-level Adversarial Augmentation (TSIAA) model for generalized medical image segmentation.
• To enhance domain-generalizable representations by exploring out-of-source data distributions.
• To improve the diversity and effectiveness of adversarial augmentation techniques.

Main Methods:

• Developed an Instance-level Image Augmenter (IIAG) using Instance-level Augmentation Modules (IAMs) based on learnable constrained Bézier transformation.
• Implemented Teacher-Student (TS) learning via an adversarial approach, alternating novel image augmentation and generalized representation learning.
• Ensured consistent and generalized characteristics between original and augmented features through continuous student and teacher updates.

Main Results:

• The proposed TSIAA model significantly improves performance in four challenging SDG tasks.
• Instance-level augmentation breaks augmentation uniformity, offering greater diversity compared to image-level methods.
• TSIAA effectively derives domain-generalizable representations by exploring out-of-source data distributions.

Conclusions:

• TSIAA offers a more diverse and effective approach to adversarial augmentation for medical image segmentation.
• The model achieves state-of-the-art results in single-source domain generalization tasks.
• This work advances the field of generalized medical image segmentation through innovative adversarial augmentation strategies.