Fast and efficient retinal blood vessel segmentation method based on deep learning network.

Area of Science:

• Ophthalmology and Medical Imaging
• Computer Vision and Artificial Intelligence

Background:

• Accurate segmentation of the retinal vascular tree is crucial for diagnosing ocular pathologies.
• Existing deep learning (DL) methods offer high accuracy but suffer from significant computational complexity and processing times.
• Clinical applications require methods that balance high segmentation performance with reduced execution times.

Purpose of the Study:

• To develop a novel deep learning (DL) based method for efficient and accurate retinal vessel tree segmentation.
• To reduce the computational complexity and processing time of retinal vessel segmentation without compromising accuracy.
• To enhance the trade-off between detection rate and detection time for automated retinal image analysis.

Main Methods:

• A new U-form deep learning (DL) architecture incorporating lightweight convolution blocks was designed.
• Specific preprocessing and data augmentation techniques tailored for retinal images and blood vessel characteristics were implemented.
• The proposed method was evaluated on the publicly available DRIVE and STARE datasets.

Main Results:

• The novel DL method achieved high segmentation accuracy, with average accuracies of 0.978 on DRIVE and 0.98 on STARE.
• The method demonstrated significantly reduced processing times, averaging 0.59 seconds per image on DRIVE and 0.48 seconds on STARE using NVIDIA GTX 980 GPUs.
• The proposed approach offers an improved balance between retinal blood vessel detection performance and computational efficiency.

Conclusions:

• The developed U-form DL architecture with lightweight convolutions provides an effective solution for high-performance, time-efficient retinal vessel segmentation.
• The tailored preprocessing and data augmentation steps contribute to the robustness and accuracy of the segmentation method.
• This approach holds promise for accelerating the clinical diagnosis of ocular diseases through automated retinal image analysis.