Related Experiment Video
Updated: Jun 15, 2025
07:23
Retinal Vascular Reactivity as Assessed by Optical Coherence Tomography Angiography
Published on: March 26, 2020
7.5K
Skeleton-guided multi-scale dual-coordinate attention aggregation network for retinal blood vessel segmentation
Wei Zhou1, Xiaorui Wang1, Xuekun Yang1
1College of Computer Science, Shenyang Aerospace University, Shenyang, China.
Computers in Biology and Medicine
|August 23, 2024
Summary
This study introduces the Skeleton-guided Multi-scale Dual-coordinate Attention Aggregation (SMDAA) network for improved retinal blood vessel segmentation. The SMDAA network effectively addresses class imbalance and image quality issues in fundus images for better medical diagnosis.
Area of Science:
- Ophthalmology
- Medical Imaging
- Computer Vision
Background:
- Deep learning is crucial for retinal blood vessel segmentation in medical diagnosis.
- Current methods struggle with class imbalance (minimal thin vessels) and poor image quality (blurred edges) in fundus images.
- These limitations lead to segmentation inaccuracies like discontinuities in vascular structures.
Purpose of the Study:
- To propose an advanced deep learning network, the Skeleton-guided Multi-scale Dual-coordinate Attention Aggregation (SMDAA) network.
- To enhance the accuracy and robustness of retinal blood vessel segmentation.
- To overcome challenges of class imbalance and image quality degradation in fundus imaging.
Main Methods:
- The proposed SMDAA network integrates three novel modules: Dual-coordinate Attention (DCA), Unbalanced Pixel Amplifier (UPA), and Vessel Skeleton Guidance (VSG).
- DCA analyzes vessel structures across multiple scales using Multi-scale Coordinate Feature Aggregation (MCFA) and Scale Coordinate Attention Decoding (SCAD).
- UPA addresses class imbalance by amplifying attention to misclassified pixels, while VSG preserves vessel continuity using anatomical information. A Feature-level Contrast (FCL) loss is also introduced.
Main Results:
- The SMDAA network demonstrated superior performance compared to existing methods across three public datasets (DRIVE, STARE, CHASE_DB1).
- The proposed modules effectively improved segmentation accuracy, particularly for thin vessels and continuous vascular structures.
- The network showed robustness against poor image quality and blurred vessel edges.
Conclusions:
- The SMDAA network offers a significant advancement in retinal blood vessel segmentation.
- It effectively tackles key challenges in fundus image analysis, leading to more reliable medical diagnoses.
- The developed approach holds promise for clinical applications requiring precise vascular segmentation.
