An improved U-net based retinal vessel image segmentation method.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Science

Background:

• Diabetic retinopathy is a leading cause of adult blindness and a common diabetes complication.
• Manual screening of retinal fundus images for diabetic retinopathy is challenging at scale.
• Accurate segmentation of retinal vessels is crucial for diagnosing diabetic retinopathy.

Purpose of the Study:

• To develop and evaluate an improved U-net network for automated retinal vessel segmentation.
• To enhance the accuracy and efficiency of diabetic retinopathy screening using deep learning.
• To address the limitations of manual analysis in large-scale retinal health screening.

Main Methods:

• Pre-processing techniques including grayscale transformation, normalization, CLAHE, and gamma transformation were applied.
• Data augmentation was utilized to prevent overfitting during the training of the neural network.
• An improved U-net architecture was proposed, integrating the Bi-FPN network for enhanced feature fusion.

Main Results:

• The proposed method achieved a vessel segmentation accuracy (ACC) of 0.9651.
• High performance metrics were obtained, including Sensitivity (SE) of 0.9767 and Area Under the Curve (AUC) of 0.9787.
• The model demonstrated strong capability in detecting vessel abnormalities, with a Specificity (SP) of 0.8604.

Conclusions:

• The improved U-net network with Bi-FPN fusion offers a robust solution for retinal vessel segmentation.
• This automated approach can significantly aid in the early detection and management of diabetic retinopathy.
• The method shows promise for large-scale, efficient retinal health screening, potentially reducing blindness rates.