Retinal Vessel Segmentation by a Transformer-U-Net Hybrid Model With Dual-Path Decoder.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate retinal vessel segmentation is crucial for diagnosing eye diseases.
• Current methods often require extensive pixel-wise annotations, increasing labor costs.
• Developing efficient segmentation frameworks with reduced annotation burden is essential.

Purpose of the Study:

• To introduce TSNet, an effective and efficient framework for retinal vessel segmentation.
• To enable a semi-supervised learning scheme utilizing rough skeleton annotations.
• To improve the balance of feature learning through multi-task outputs.

Main Methods:

• Designed a Transformer-CNN hybrid model integrating a Transformer module within a U-Net architecture for long-range interaction capture.
• Implemented a dual-path decoder in the U-Net for multi-task outputs, including auxiliary vessel skeleton prediction.
• Developed a skeleton semi-supervised learning scheme using a mean teacher model for pseudo-annotation generation and correction.

Main Results:

• TSNet achieved state-of-the-art performance on retinal vessel segmentation across five benchmarking datasets.
• The framework demonstrated effectiveness in a fully supervised setting.
• The proposed semi-supervised scheme significantly reduced annotation efforts while maintaining high performance.

Conclusions:

• TSNet offers a powerful and efficient solution for retinal vessel segmentation.
• The novel annotation process and semi-supervised learning scheme lower the barrier for practical application.
• This approach facilitates more accessible and efficient eye disease diagnosis through improved medical image analysis.