HDC-Net: A hierarchical dilation convolutional network for retinal vessel segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Ophthalmic diseases often manifest through abnormalities in retinal blood vessels.
• Early detection and prevention of these diseases are crucial for patient outcomes.
• Automated analysis of retinal vasculature using deep learning offers significant potential.

Purpose of the Study:

• To develop an advanced deep learning model for precise retinal vessel segmentation.
• To improve the accuracy and efficiency of identifying morphological and structural details of retinal blood vessels.

Main Methods:

• Proposed a hierarchical dilation convolutional network (HDC-Net) for pixel-to-pixel retinal vessel extraction.
• Incorporated a hierarchical dilation convolution (HDC) module to capture subtle blood vessels.
• Utilized an improved residual dual efficient channel attention (RDECA) module for enhanced feature discrimination.
• Implemented structured Dropblock to mitigate network overfitting.

Main Results:

• HDC-Net demonstrated superior segmentation performance across three benchmark datasets (DRIVE, CHASE-DB1, STARE).
• Achieved high sensitivity (e.g., 0.8369), specificity (e.g., 0.9866), accuracy (e.g., 0.9751), F1-score (e.g., 0.8385), and AUC (e.g., 0.9913).
• Outperformed existing state-of-the-art retinal vessel segmentation models.

Conclusions:

• HDC-Net effectively and accurately performs retinal vessel segmentation.
• The model's ability to capture fragile vessels and enhance feature discrimination contributes to its high performance.
• This technology holds promise for early diagnosis and proactive management of ophthalmic conditions.