Purchase this article with an account.
Tristan Hormel, Jie Wang, Yukun Guo, Thomas S Hwang, David Huang, Yali Jia; A Comparison of OCT Angiography Non-Perfusion Area Measurements Made With a Rules-Based Approach and Artificial Intelligence. Invest. Ophthalmol. Vis. Sci. 2022;63(7):1158.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To compare optical coherence tomography angiography (OCTA) non-perfusion area (NPA) measurements made using artificial intelligence (AI) and a traditional rules-based approach.
NPA measurements were performed on eyes with diabetic retinopathy (DR), with severity determined by a central reading center based on 7-field color fundus photography using the early treatment of diabetic retinopathy (ETDRS) scale (referable ≥ 35; vision-threatening ≥ 53). We assessed the accuracy of DR diagnosis based on NPA measurements made with each approach by calculating the area under receiver operating characteristic curve, and the robustness of each method by measuring repeatability and the effect of signal strength. Measurements were assessed using 959 OCTA scans of eyes diagnosed with DR (2 non-referable; 670 referable/non-vision threatening; 287 vision threatening). The source of the data is the DRCR Retina Network, but the analyses, content and conclusions presented herein are solely the responsibility of the authors and have not been reviewed or approved by DRCR Retina Network. 15 repeat scans in this dataset were used to test repeatability. Since the DRCR dataset did not contain healthy controls, 128 healthy eyes imaged at Oregon Health & Science University (OHSU) were added to assess NPA-based diagnostic accuracy. The two methods considered here are 1) a rules-based approach that determined NPA using a vessel distance map, and 2) an AI-based approach that used a U-net-like architecture (Fig. 1). Both approaches quantified NPA in the superficial vascular complex and inner retinal slab. The AI model was trained with a separate dataset from OHSU containing 492 scans of eyes with varying degrees of DR.
The AI approach had better diagnostic performance than the rules-based method for referable vs. non-referable DR (Fig. 2). The AI-based NPA measurements were also less correlated with signal strength in the SVC (R2 = 0.121 rules-based vs. R2 = 0.002 AI-based). The rules based approach achieved slightly higher repeatability in the SVC (intraclass correlation coefficient = 0.945 rules-based vs. = 0.923 AI-based methods).
NPA measurements made on a clinical dataset using an AI-based approach have a higher diagnostic accuracy for diagnosing referable DR and are more robust with respect to scan quality than NPA measurements from a rules-based approach.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
This PDF is available to Subscribers Only