Automated method for quantitative analysis of iris fluorescein angiography based on machine learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy is a primary cause of vision loss, frequently leading to neovascular glaucoma.
• Early identification of neovascularisation of the iris (NVI) is critical for timely treatment.
• Conventional methods for NVI detection have limitations in identifying early signs.

Purpose of the Study:

• To develop and evaluate a deep learning-based automated system for detecting and quantifying peripupillary leakage in iris fluorescein angiography (IFA) images.
• To assess the system's performance against manual annotations by clinical experts.

Main Methods:

• A YOLOv8n-based segmentation model was trained on 2,449 IFA images for pupil localization.
• A leakage circularity detection algorithm was developed to quantify peripupillary fluorescein leakage.
• Performance was evaluated on 131 IFA images using metrics like MAE, MAPE, and IoU, comparing results with two clinical experts.

Main Results:

• The automated method showed significantly lower Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) compared to manual expert assessments.
• The algorithm achieved a 39.3% Intersection over Union (IoU), indicating high segmentation accuracy with minor spatial misalignment.
• Inter-clinician agreement (IoU of 54.8%) highlighted inherent subjectivity in human evaluations.

Conclusions:

• The deep learning approach offers superior consistency and accuracy in quantifying peripupillary leakage over manual methods.
• This automated system can reduce variability and subjectivity in NVI diagnosis.
• The technology holds potential for earlier NVI detection, enhanced clinical workflow, and improved diabetic retinopathy management.