CS-UNet: Cross-scale U-Net with Semantic-position dependencies for retinal vessel segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Accurate retinal vessel segmentation is crucial for diagnosing and treating eye conditions.
• Existing methods like U-Net struggle with complex retinal images featuring small vessels, low contrast, and lesions.
• Limitations in global modeling affect the performance of pure convolutional networks.

Purpose of the Study:

• To introduce a novel deep learning model, Cross-scale U-Net with Semantic-position Dependencies (CS-UNet), for enhanced retinal vessel segmentation.
• To address the limitations of traditional U-Net in capturing global context and cross-scale information.

Main Methods:

• Developed a Semantic-position Dependencies Aggregator (SPDA) to integrate semantic and positional relationships for global context.
• Introduced a Cross-scale Relation Refine Module (CSRR) to facilitate dynamic cross-scale information interaction.
• Incorporated SPDA into the encoder layers and CSRR to guide the up-sampling process.

Main Results:

• CS-UNet demonstrated superior performance in retinal vessel segmentation compared to existing state-of-the-art methods.
• Evaluated on three public datasets: DRIVE, CHASE_DB1, and STARE, confirming robust performance.
• The proposed modules effectively improved the model's ability to handle intricate vessel structures and low-contrast regions.

Conclusions:

• CS-UNet offers a significant advancement in retinal vessel segmentation accuracy.
• The integration of semantic-position dependencies and cross-scale interaction modules enhances the model's effectiveness.
• This approach holds promise for improving the early diagnosis and management of retina-related diseases.