Rethinking Data Augmentation for Single-Source Domain Generalization in OCT Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Domain shifts between instruments pose challenges for Optical Coherence Tomography (OCT) image segmentation.
• Variations in imaging devices and clinical centers lead to differences in OCT data acquisition.

Purpose of the Study:

• To present a novel Style and Structure Data Augmentation (SSDA) method to enhance OCT segmentation model adaptability.
• To address domain shifts caused by stylistic and structural variations in OCT images.

Main Methods:

• Developed SSDA incorporating modality-specific NURBS curves for style enhancement.
• Implemented global and local elastic deformations to simulate retinal curvature and layer-specific changes.
• Validated SSDA through single-domain generalization experiments on five diverse OCT datasets.

Main Results:

• SSDA demonstrated superior performance over existing methods in cross-domain OCT segmentation.
• Achieved approximately 1.6% higher Dice and 2.6% improved MIOU across five generalization experiments.
• Highlighted robust generalization capabilities on unseen OCT domains from different sources.

Conclusions:

• SSDA effectively mitigates domain shift challenges in OCT image segmentation.
• The proposed method improves the adaptability and accuracy of segmentation models across diverse OCT datasets.
• SSDA offers a promising solution for reliable OCT image analysis in multi-center clinical settings.