Diabetic and Hypertensive Retinopathy Screening in Fundus Images Using Artificially Intelligent Shallow Architectures

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Science

Background:

• Retinal blood vessels are key biomarkers for diabetic retinopathy and hypertensive retinopathy.
• Manual segmentation of retinal vasculature is time-consuming and prone to error.
• Existing automatic methods are often computationally intensive and lack robustness.

Purpose of the Study:

• To propose novel, efficient deep learning architectures for accurate retinal vasculature segmentation.
• To enable automated screening for diabetic and hypertensive retinopathies using color fundus images.
• To improve upon the robustness and computational efficiency of current segmentation techniques.

Main Methods:

• Development of two shallow deep learning architectures: dual-stream fusion network (DSF-Net) and dual-stream aggregation network (DSA-Net).
• Application of semantic segmentation on raw color fundus images.
• Performance evaluation on three public datasets: DRIVE, STARE, and CHASE-DB1.

Main Results:

• DSF-Net and DSA-Net achieved high segmentation performance across all datasets.
• Specific metrics include accuracy (up to 97.25%), sensitivity (up to 86.07%), specificity (up to 98.38%), and AUC (up to 98.65%).
• DSA-Net demonstrated superior sensitivity, effectively detecting smaller vessels and minimizing false negatives.

Conclusions:

• The proposed DSF-Net and DSA-Net offer accurate and efficient automated retinal vasculature segmentation.
• These models facilitate early detection and differentiation of diabetic and hypertensive retinopathies.
• The generated segmentation masks can aid in analyzing vessel morphology for various retinal disorders.