Abstract
BACKGROUND: Diabetic retinopathy is a leading cause of vision impairment, often progressing to neovascular glaucoma. Early detection of neovascularisation of the iris (NVI) is crucial for timely intervention. Traditional diagnostic methods, such as slit-lamp examination, have limitations in identifying early-stage NVI. This study presents a deep learning-based automated approach for analysing iris fluorescein angiography (IFA) images to detect and quantify peripupillary leakage, a key indicator of NVI. METHODS: A dataset of 2,449 IFA images was used to train a YOLOv8n-based segmentation model for precise pupil localisation. A leakage circularity detection algorithm was developed to quantify peripupillary fluorescein leakage. The algorithm's performance was evaluated using an independent test set of 131 clinically standardized IFA images. Performance metrics included mean absolute error (MAE), mean absolute percentage error (MAPE), and intersection over union (IoU). Results were compared with manual annotations from two clinical experts. RESULTS: The proposed method demonstrated a significant reduction in MAE (20.81 degrees) and MAPE (21.64%) compared to Clinical Staff 1 (MAE: 34.23 degrees, MAPE: 58.38%) and Clinical Staff 2 (MAE: 43.17 degrees, MAPE: 75.71%). The algorithm achieved an IoU of 39.3%, slightly lower than Clinical Staff 1 (44.5%) and Clinical Staff 2 (41.7%), indicating high segmentation accuracy but minor spatial misalignment. The inter-clinician agreement yielded an IoU of 54.8%, highlighting subjectivity in human assessments. CONCLUSIONS: The deep learning-based approach provides superior consistency and accuracy in quantifying peripupillary fluorescein leakage compared to manual expert annotations. While human experts demonstrated slightly higher spatial precision, the algorithm significantly reduces variability and subjectivity in leakage quantification. This automated method has the potential to enhance early detection of NVI, improve clinical workflow efficiency, and assist ophthalmologists in diagnosing DR. Further optimization will focus on refining spatial segmentation accuracy.