Robust semi-automatic vessel tracing in the human retinal image by an instance segmentation neural network.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computational Biology

Background:

• Vascular morphology and hierarchy are crucial for effective blood perfusion.
• The human retinal circulation is a complex network originating from the optic nerve head (ONH).
• Accurate tracing of retinal vascular branching is vital for morphological quantification but technically challenging.

Purpose of the Study:

• To develop a robust semi-automatic algorithm for tracing individual vascular trees in human retinal images.
• To enable detailed morphological analysis of retinal vasculature and its hierarchy.
• To improve the accuracy and robustness of retinal vessel tracing for disease-related studies.

Main Methods:

• Utilized an instance segmentation neural network (InSegNN) to separate and label individual vascular trees.
• Implemented pseudotemporal learning, spatial multisampling, and dynamic probability map strategies to enhance accuracy.
• Applied the algorithm to human fundus images for vessel tracing.

Main Results:

• Achieved 83% specificity and 91% precision with 71% sensitivity.
• Demonstrated a 50% improvement in symmetric best dice (SBD) compared to existing literature.
• Outperformed the baseline U-net model in vessel tracing accuracy.
• Successfully traced individual vessel trees while retaining hierarchical information.

Conclusions:

• The InSegNN algorithm provides a robust method for tracing individual retinal vascular trees and their hierarchy from fundus images.
• This technique facilitates detailed vascular morphology analysis relevant to retinal diseases.
• InSegNN offers a significant advancement for research in ophthalmology and medical imaging of the retinal vasculature.