Retinal vascular junction detection and classification via deep neural networks.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal vascular trees exhibit complex structures with intersecting and overlapping vessels.
• Vascular junctions, including bifurcations and crossovers, are crucial for analyzing retinal vascular diseases and morphology.
• Accurate detection and classification of these junctions are essential for clinical applications.

Purpose of the Study:

• To propose a novel two-stage pipeline for detecting and classifying retinal vascular junctions directly from color retinal images.
• To improve the accuracy and efficiency of identifying vascular bifurcations and crossovers in retinal images.
• To provide a method that does not require vessel segmentation or skeletonization preprocessing.

Main Methods:

• A RCNN-based Junction Proposal Network detects potential junction locations.
• A Junction Refinement Network eliminates false detections.
• A Junction Classification Network, sharing the same structure as the refinement network, classifies junctions as bifurcation or crossover.

Main Results:

• The proposed approach achieved 70% F1-score on the DRIVE dataset and 60% F1-score on the IOSTAR dataset.
• Performance surpassed state-of-the-art methods by 4.5% (DRIVE) and 1.7% (IOSTAR).
• The method demonstrated high and balanced precision and recall values.

Conclusions:

• A new junction detection and classification method for retinal images was successfully developed.
• The approach operates directly on color retinal images, eliminating the need for vessel segmentation or skeleton preprocessing.
• Superior performance validates the effectiveness of the proposed junction analysis method.