Retinal vessel segmentation driven by structure prior tokens.

Area of Science:

• Ophthalmic imaging analysis
• Medical image segmentation
• Biomedical engineering

Background:

• Accurate retinal vessel segmentation in OCTA images is crucial for diagnosing eye diseases like diabetic retinopathy.
• Traditional methods struggle with complex vascular structures and local features, leading to suboptimal accuracy.
• Existing techniques often neglect the intrinsic structural properties of retinal vessels.

Purpose of the Study:

• To integrate structural priors, encoding vessel morphology and topology, into a segmentation framework.
• To enhance the accuracy and robustness of retinal vessel segmentation, especially in challenging image regions.
• To improve the preservation of vascular integrity and continuity in OCTA images.

Main Methods:

• A generative image segmentation framework utilizing a latent embedding space for retinal vessel priors.
• A prior-driven network that learns vessel priors from ground truth data and stores them in a codebook.
• Encoding semantic features of OCTA images with learned prior tokens for vessel reconstruction.

Main Results:

• The proposed network demonstrated superior performance over state-of-the-art methods on three OCTA datasets (ROSE-1, ROSE-2, OCTA-Z).
• Achieved average Dice scores of 77.63% (ROSE-1), 71.01% (ROSE-2), and 81.11% (OCTA-Z).
• Qualitative and quantitative evaluations confirmed the network's effectiveness in preserving vascular structure integrity.

Conclusions:

• The developed network successfully learned and applied implicit vessel priors for improved OCTA segmentation.
• The latent prior token reconstruction approach offers a promising solution for retinal vessel pattern representation.
• Future work includes extending the method for broader retinal structure segmentation and disease classification.