FRD-Net: a full-resolution dilated convolution network for retinal vessel segmentation

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate retinal vessel segmentation is crucial for diagnosing and planning surgeries for retinal diseases.
• Conventional U-Net architectures struggle with fine, low-contrast vessels due to resolution loss during encoding and decoding.
• Existing methods often fail to preserve fine details and context necessary for precise segmentation.

Purpose of the Study:

• To introduce FRD-Net, an effective full-resolution network for enhanced retinal vessel segmentation.
• To overcome the limitations of conventional networks in segmenting challenging retinal vasculature.
• To improve the accuracy and generalization of automated retinal vessel segmentation.

Main Methods:

• FRD-Net utilizes a backbone network with multi-resolution dilated convolutions for horizontal and vertical expansion, preserving full image resolution.
• Dilated residual modules integrate multi-scale features, enabling continuous learning of contextual information.
• A multi-scale feature fusion module (MFFM) fuses deep features with the original image to recover edge details.

Main Results:

• FRD-Net demonstrates superior performance and generalization capabilities compared to state-of-the-art segmentation algorithms on multiple datasets.
• The proposed network achieves accurate segmentation of fine and low-contrast retinal blood vessels.
• FRD-Net requires fewer model parameters while maintaining high accuracy.

Conclusions:

• FRD-Net offers a significant advancement in automated retinal vessel segmentation.
• The full-resolution approach effectively addresses the limitations of conventional encoder-decoder networks.
• FRD-Net shows great potential for clinical applications in ophthalmology and retinal disease management.