Retinal vessel segmentation based on multi-scale feature and style transfer.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal vessel segmentation is crucial for diagnosing eye diseases.
• Current deep learning methods struggle with cross-domain data and small vessel segmentation.
• Overly complex models present challenges in practical application.

Purpose of the Study:

• To propose a novel, lightweight network (MSFST-NET) for improved retinal vessel segmentation.
• To enhance the model's ability to handle cross-domain datasets and segment small blood vessels.
• To avoid overly complex model architectures.

Main Methods:

• Developed MSF-Net with a selective kernel (SK) module for multi-scale feature extraction.
• Introduced a style transfer module to reduce domain discrepancies.
• Implemented a pseudo-label learning strategy to boost generalization.

Main Results:

• MSFST-NET demonstrated improved segmentation of small blood vessels.
• The style transfer and pseudo-labeling effectively improved cross-domain performance.
• Experiments on DRIVE and CHASE_DB1 datasets showed superior results compared to state-of-the-art methods.

Conclusions:

• MSFST-NET offers an effective solution for retinal vessel segmentation, particularly for cross-domain challenges.
• The proposed methods enhance model generalization and segmentation accuracy.
• This approach provides a more robust and efficient tool for clinical applications.