Hallucinated domain generalization network with domain-aware dynamic representation for medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation models struggle with performance degradation across different acquisition protocols and domains.
• This degradation is often caused by overfitting to source data and poor adaptability to new target domains.

Purpose of the Study:

• To develop a novel network for medical image segmentation that enhances generalization across unseen domains.
• To address overfitting and improve dynamic adaptability in segmentation models.

Main Methods:

• Proposed a hallucinated domain generalization network with domain-aware dynamic representation.
• Introduced an uncertainty-aware dynamic hallucination module using Bézier curves and uncertainty-aware offset for synthetic image generation.
• Developed a domain-aware dynamic representation module for mapping input features to a unified source domain space using style prototypes and similarity weights.

Main Results:

• The proposed method effectively breaks source domain limitations while preserving anatomical structures.
• It alleviates model overfitting to specific source domain styles.
• Experiments on fundus and prostate MRI datasets showed superior performance compared to state-of-the-art methods.

Conclusions:

• The novel "hallucination during training, dynamic representation during testing" scheme significantly improves generalization for medical image segmentation.
• The approach effectively mitigates performance degradation caused by domain shift.
• This method offers a robust solution for cross-domain medical image segmentation tasks.