Related Experiment Video
Updated: Jun 30, 2025
Retinal Vascular Reactivity as Assessed by Optical Coherence Tomography Angiography
Published on: March 26, 2020
HRD-Net: High resolution segmentation network with adaptive learning ability of retinal vessel features
Jianhua Liu1, Dongxin Zhao2, Juncai Shen3
1School of Electrical and Electronic Engineering, Shijiazhuang Tiedao University, Shijiazhuang, 050043, China; Hebei Provincial Collaborative Innovation Center of Transportation Power Grid Intelligent Integration Technology and Equipment, School of Electrical and Electronic Engineering, Shijiazhuang Tiedao University, Shijiazhuang, China.
HRD-Net, a novel high-resolution network, enhances retinal blood vessel segmentation using Deformable Convolution v3. It accurately captures fine vascular details, improving early disease detection with fewer parameters.
Area of Science:
- Medical Imaging
- Computer Vision
- Biomedical Engineering
Background:
- Retinal blood vessel segmentation is vital for diagnosing health conditions.
- Standard convolutional networks struggle with complex vascular structures and spatial information loss.
- Existing methods often result in over-segmentation or missed vessels.
Purpose of the Study:
- To develop an advanced network for accurate retinal blood vessel segmentation.
- To address limitations of conventional methods in capturing fine vascular details and preserving spatial information.
- To improve early detection of health issues through enhanced retinal image analysis.
Main Methods:
- Developed HRD-Net, a high-resolution network incorporating Deformable Convolution v3.
- Implemented a feature enhancement cascade module for flexible adaptation to vessel morphology.
- Utilized a global aggregation module for fusing semantic and spatial information at full resolution.
- Optimized network architecture and activation/normalization methods for efficiency.
Main Results:
- HRD-Net demonstrated superior performance on DRIVE, STARE, and CHASE_DB1 datasets.
- Achieved improved segmentation accuracy across multiple metrics including F1, ACC, SE, SP, AUC, and IOU.
- Showcased effectiveness in segmenting fine blood vessels and maintaining vessel continuity.
- Outperformed existing methods with a reduced parameter count.
Conclusions:
- HRD-Net offers a significant advancement in retinal blood vessel segmentation.
- The proposed architecture effectively handles complex vascular structures and preserves crucial spatial details.
- This method holds promise for more accurate and efficient early detection of eye diseases and systemic health conditions.
