DG-TTA: Out-of-Domain Medical Image Segmentation Through Augmentation, Descriptor-Driven Domain Generalization, and...

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Pre-trained medical deep learning segmentation models struggle with out-of-domain images.
• Poor segmentation quality hinders clinical applications of AI in medical imaging.

Purpose of the Study:

• To develop a robust method for domain-generalized pre-training and test-time adaptation in medical image segmentation.
• To achieve high-quality segmentation performance on unseen medical imaging domains.

Main Methods:

• Utilized a robust generalizing descriptor (SSC) and intensity augmentation (GIN) for domain-generalized pre-training.
• Implemented test-time adaptation using a consistency scheme with augmentation-descriptor combination for unseen scans.
• Evaluated performance on five public datasets (3D CT and MRI) across abdominal, spine, and cardiac imaging.

Main Results:

• Significant improvements in segmentation performance across cross-domain scenarios (CT to MRI).
• Achieved substantial Dice score increases in abdominal (+46.2, +28.2), spine (+72.9), and cardiac (+14.2, +55.7) imaging (p < 0.001).
• Demonstrated effective bridging of domain gaps with a compact and efficient methodology.

Conclusions:

• The proposed method enables high-quality medical image segmentation in unseen domains.
• Domain-generalized pre-training and test-time adaptation significantly enhance model generalization.
• The approach allows optimal and independent use of source and target data.