Parallel Architecture of Fully Convolved Neural Network for Retinal Vessel Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate retinal blood vessel extraction is crucial for diagnosing various retinal diseases.
• Existing segmentation methods face challenges in accuracy and efficiency.
• Public datasets like DRIVE and STARE are vital for evaluating retinal vessel segmentation algorithms.

Purpose of the Study:

• To propose a parallel fully convolved neural network architecture for enhanced retinal blood vessel segmentation.
• To evaluate the impact of image preprocessing on network performance.
• To demonstrate the method's effectiveness and efficiency on standard datasets.

Main Methods:

• A parallel fully convolved neural network architecture was developed for retinal blood vessel segmentation.
• The network's performance was assessed using different image preprocessing techniques.
• Experiments were conducted on the DRIVE and STARE public retinal image databases.

Main Results:

• The proposed method achieved high performance metrics: 96.37% accuracy, 86.53% sensitivity, and 98.18% specificity.
• The architecture demonstrated data independence by successfully segmenting abnormal STARE images using a DRIVE-trained model.
• Vessel extraction performance was consistent across varying vessel thicknesses.

Conclusions:

• The developed neural network architecture offers a robust and accurate solution for retinal blood vessel segmentation.
• The method outperforms existing segmentation techniques and is suitable for real-time applications due to CPU implementation and low computational cost.
• The system achieves rapid image processing, extracting vessels in less than 2 seconds per image.