Imaging Features of Vessels and Leakage Patterns Predict Extended Interval Aflibercept Dosing Using Ultra-Widefield Angiography in Retinal Vascular Disease: Findings From the PERMEATE Study

Abstract

Diabetic Macular Edema (DME) and macular edema secondary to retinal occlusion (RVO) are the two most common retinal vascular causes of visual impairment and leading cause of worldwide vision loss. The blood-retinal barrier is the key barrier for maintaining fluid balance within the retinal tissue. Vascular Endothelial Growth Factor (VEGF) has a significant role in the permeability of the blood-retinal barrier, which also leads to appearance of leakage foci. Intravitreal anti-VEGF therapy is the current gold standard treatment and has been demonstrated to improve macular thickening, improve vision acuity and reduce vascular leakage. However, treatment response and required dosing interval can vary widely across patients. Given the role of the blood-retinal barrier and vascular leakage in the pathogenesis of these disorders, the goal of this study was to present and evaluate new computer extracted features relating to morphology, spatial architecture and tortuosity of vessels and leakages from baseline ultra-widefield fluorescein angiography (UWFA) images. Specifically, we sought to evaluate the role of these computer extracted features from baseline UWFA images. Notably, these UWFA images were obtained from IRB-approved PERMEATE clinical trial [1], [2] to distinguish eyes tolerating extended dosing intervals (n = 16) who are referred to as non-rebounders and those who require more frequent dosing (n = 12) and are called rebounders based on visual acuity loss with extended dosing challenges. A total of 64 features encapsulating different morphological and geometrical attributes of leakage patches including the anatomical (shape, size, density, area, minor and major axis, orientation, area, extent ratio, perimeter, radii) and geometrical characteristics (the proximity of each leakage foci to main vessels, to other leakage foci and to optical disc) as well as 54 tortuosity features (tortuosity of whole vessel network, local tortuosity of vessels in the vicinity of leakage foci) were extracted. The most significant and predictive biomarkers related to treatment response were proximity of leakage nodes to major and minor eye vessels as well as local vasculature tortuosity in the vicinity of the leakages. The imaging features were then used in conjunction with a Linear Discriminant Analysis (LDA) classifier to distinguish rebounders from non-rebounders. The 3-fold cross-validated Area Under Curve (AUC) was found to be 0.82 for the morphological based features and 0.85 for the tortuosity based features. Our findings suggest higher variation in leakage node proximity to retinal vessels in eyes tolerating extended interval dosing. In contrast, eyes with increased local vascular tortuosity demonstrated less tolerance of increased dosing interval. Moreover, a class activation map generated by a deep learning model identified regions that corresponded to regions of leakages proximal to the vessels, providing confirmation of the validity of predictive image features extracted from these regions in this study.

Fig. 1:

Flowchart illustrating the steps involved for extracting and evaluating the UWFA imaging features from the PERMEATE trial.

Fig. 2:

Illustration of the process of calculating the distance of leakage foci to main vessels. (a) Fused images of segmented leakage foci and corresponding vasculature network. (b) The bounding boxes are defined around each leakage foci (c) The proximity of leakage foci to proximal primary vessels within the bounding box is calculated.

Fig. 3:

The process of calculating the tortuosity of vessels around leakage foci. (a) Bounding boxes (yellow) are defined around each leakage focus (red dots) on the segmented vessel network (green). (b) The tortuosity of vessels around the leakage foci is calculated.

Fig. 4:

a) The clustergram of morphological and geometrical based features. The Y label represents the individual patients (The rebounders are specified with 0s and non-rebounders are specified with 1s). The X-axis reflects the reduced-dimension features (ten features selected using PCA). b) The boxplot of median of variance of distance of leakage foci in the vicinity of main vessels for the rebounders and non-rebounders (n=28). The feature is identified as the most significant feature during the feature selection process.

Fig. 5:

a) The clustergram of tortuosity based features The Y label represents the individual patients (The rebounders are specified with 0s and non-rebounders are specified with 1s.) The X-axis reflects the reduced-dimension features (ten features selected using PCA). b) The boxplot of median of maximum tortuosity of vessels in locality of leakage foci in the 2 cohorts (n=28). which was identified as the most significant feature during the feature selection process.

Fig. 6:

Visualizing the deep learning model in order to recognize the important regions of image for the classifier. The gradients of the last layer are shown with respect to the previous convolutional layer.