Automated Quantitative Ultra-widefield Angiographic Leakage Assessment Demonstrates Independent Replication of PRIME Diabetic Retinopathy Clinical Trial

Abstract

Purpose : Ultra-widefield fluorescein angiography (UWFA) is an imaging technique that allows near panretinal visualization of retinal vascular dynamics, such as leakage. Quantitative assessment of leakage has historically been challenging and laborious. The purpose of the analysis is to assess the feasibility of using an automated deep learning (DL) model to replicate the UWFA leakage outcomes from the reader-based assessments from the PRIME diabetic retinopathy clinical trial.

Methods : PRIME was an IRB-approved clinical trial that sought to determine the effect of panretinal leakage index (PLI) guided intravitreal aflibercept injections (IAI) vs real-time DR severity scale (DRSS) guided IAI using monthly UWFA on PLI (among other clinical outcomes). In the PRIME study, PLI was determined by an independent system for leakage quantification with certified reader review and correction of segmentation errors as needed. For this analysis, we used an automated DL leakage segmentation system that was trained on 1439 scans from 384 patients in TensorFlow with the TransUNET model architecture. PLI was calculated by dividing leakage area by region of interest area.

Results : DL segmentation produced a statistically equivalent mean total PLI to the original trial at 13 of 14 time points. For the final time point, the model was under-sensitive due to a change in image quality, and a reduction in the confidence threshold resulted the in replication of the last time point. Our model demonstrated that the DRSS-guided arm (n = 15) showed a reduction in PLI of 28.3 % (p = 0.34) while the PLI guided arm (n = 16) showed a reduction of 57.1% (p = 0.003) from week 0 to week 52 (Figure 1). This outcome is equivalent to the original trial which showed the DRSS-guided arm had a reduction of 24.6% (p=0.49) and the PLI-guided arm had a reduction of 54.6% (p < 0.0001) (Figure 2).

Conclusions : The automated UWFA leakage model displayed promising clinical utility based on quantitative performance measures and qualitative inspection of segmentation overlays. Additional research will include optimization of automated thresholding to facilitate more versatile assessments.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Automated segmentation of PLI produces statistically equivalent mean PLI to the original trial at all 14 time points.

 

Original PRIME clinical trial outcomes demonstrating monthly PLI levels.