Unsupervised domain adaptation teacher-student network for retinal vessel segmentation via full-resolution refined...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal blood vessel morphology changes are indicators of systemic diseases like hypertension and diabetes.
• Automated segmentation of retinal blood vessels in fundus images aids in early disease detection.
• Domain shift, a challenge in applying models across datasets, significantly reduces segmentation accuracy.

Purpose of the Study:

• To propose a novel unsupervised domain adaptation method for retinal blood vessel segmentation.
• To overcome the domain shift problem in automated fundus image analysis.
• To improve the robustness and accuracy of segmentation models across different datasets.

Main Methods:

• A teacher-student framework was employed to generate pseudo-labels for target domain images.
• The student network was trained using a combination of source domain loss and domain adaptation loss.
• A full-resolution refined model was reconstructed by computing training loss at multiple semantic levels and resolutions.

Main Results:

• The method demonstrated high performance on the DRIVE and STARE datasets, achieving accuracy, sensitivity, and specificity above 0.96, 0.84, and 0.97 respectively across domains.
• Outperformed most state-of-the-art unsupervised domain adaptation methods.
• Achieved the best F1 score (0.8053) from STARE to DRIVE and a competitive F1 score (0.8001) from DRIVE to STARE.

Conclusions:

• The proposed unsupervised domain adaptation method effectively addresses the domain shift problem in retinal blood vessel segmentation.
• The approach offers a robust solution for cross-dataset generalization in medical image analysis.
• This method shows significant potential for improving automated diagnosis of systemic diseases through retinal imaging.