Exploratory Assessment of Features Capable of Distinguishing Between Ischemic and Non-ischemic Candidate Regions in Ultra-widefield Fluorescein Angiography

Abstract

Purpose : Quantifying retinal ischemia is important for evaluating underlying diagnosis, disease burden and progression risk. Manual segmentation of ischemia in ultra-widefield fluorescein angiography (UWFA) is laborious and involves potential variability between human graders. To address these challenges, this study investigates features that distinguish ischemic from non-ischemic candidate regions and uses those features to classify regions as ischemic or non-ischemic utilizing machine learning.

Methods : This IRB-approved retrospective image analysis study utilized UWFA images with various levels of retinal ischemia. Images were manually segmented by expert trained graders to establish ground truth. A previously developed pipeline involving deep-learning-based UWFA blood vessel segmentation, rule-based generation of potentially ischemic candidate regions between blood vessels, and extraction of candidate region features was utilized. For purposes of this analysis, ground truth images were primarily restricted to high-quality images during the early phases of the UWFA. Intensity, shape, location, and texture features were used to classify candidate regions as ischemic or non-ischemic using a gradient boosted decision tree classifier. The model was evaluated using a 20-fold cross-validated approach.

Results : 101,469 candidate regions were segmented from 127 eyes. 13% of candidate regions were completely or partially ischemic. The model demonstrated area under the precision-recall curve (AUPRC) = 0.40, significantly better than random chance AUPRC = 0.13 (p < 0.001). Important features in the model included 10^th percentile intensity, short run low gray level emphasis, robust mean absolute deviation (all negatively associated with ischemia, p < 0.001), and skewness (positively associated with ischemia, p < 0.001).

Conclusions : This study identified features that enable classification of candidate regions as ischemic or non-ischemic with moderate performance characteristics. This feature interrogation also highlighted the complexity of the challenges of automated identification of ischemia. Future work will focus on identifying additional features to improve performance, expanding the ground truth dataset, and exploring other methods of determining ground truth and candidate region generation.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.