Two-stage adversarial learning based unsupervised domain adaptation for retinal OCT segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep learning for optical coherence tomography (OCT) segmentation enables quantitative analysis.
• Domain shift due to varying acquisition devices/protocols degrades segmentation model performance.
• Addressing domain shift is crucial for robust OCT image analysis.

Purpose of the Study:

• Propose a two-stage adversarial learning network (TSANet) for unsupervised cross-domain OCT segmentation.
• Overcome performance degradation caused by domain shift in OCT imaging.
• Enable accurate segmentation across different OCT datasets without manual relabeling.

Main Methods:

• Employ a Fourier transform approach for image-level style adaptation in the first stage.
• Utilize adversarial learning with a segmenter and discriminator for inter-domain consistency.
• Implement pseudo-labeling and fine-tuning in the second stage to enhance generalization.

Main Results:

• Achieved significant improvements in intersection over union (IoU) by 8.34%, 55.82%, and 3.53% across three test sets.
• Demonstrated superior performance compared to state-of-the-art domain adaptation methods.
• Validated effectiveness on choroid and retinoschisis segmentation tasks.

Conclusions:

• TSANet effectively achieves cross-domain generalization through multi-level adaptation strategies.
• Reduces the need for manual annotation when adapting deep learning models to new OCT data.
• Facilitates broader application of OCT segmentation models across diverse datasets.