Deep Learning-Based Choroidal Boundary Detection in Geographic Atrophy Using Spectral-Domain Optical Coherence...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Geographic atrophy (GA) affects central vision, necessitating accurate monitoring.
• Spectral-domain OCT (SD-OCT) is crucial for visualizing retinal layers, including the choroid.
• Automated choroidal boundary detection aids in analyzing GA progression.

Purpose of the Study:

• Evaluate deep learning model challenges for choroidal boundary detection in GA eyes.
• Assess the workflow efficiency of an AI-assisted manual verification approach.

Main Methods:

• Retrospective analysis of 5723 SD-OCT scans from GA patients.
• Utilized NMI ChoroidAI for segmenting choroidal inner (CIB) and outer (COB) boundaries.
• Compared AI-assisted workflow against manual segmentation for accuracy and time.

Main Results:

• CIB detection showed high accuracy (94.8% accurate, F1 score 0.97).
• COB detection was more prone to errors (19.0% deviation), but 94.2% were acceptable with minor deviations.
• AI-assisted workflow reduced processing time by ~90% (7h manual vs. 45min AI+human).

Conclusions:

• Deep learning model shows limitations in COB detection due to artifacts.
• AI-assisted approach substantially reduces human effort in choroidal segmentation.
• Mandatory human verification is essential to correct errors before clinical application.