Abstract
Purpose :
Computational image recognition and machine learning techniques have been used to develop algorithms that distinguish normal retinal morphological features from distorted contours occurring due to the presence of fluid between and within tissue layers. A deep neural network has previously demonstrated efficacy for the detection and quantification of exudative fluid in age-related macular degeneration. However, it lacks validation in diabetic macular edema (DME) or retinal vein occlusion (RVO). This study aims to externally evaluate the performance of NOA against expert graders for the classification and quantification of fluid in DME and RVO.
Methods :
Retrospective, non-randomized cohort study performed at Cole Eye Institute, Cleveland, OH. Scans were graded by the Notal OCT Analyzer (NOA, Notal Vision Ltd, Tel Aviv, Israel) machine learning and image recognition algorithm and manually measured by two expert graders who assessed the quantification of fluid using OCTExplorer (Iowa Institute for Biomedical Imaging, Iowa City, IA). The results from manual and NOA reading were compared to determine interclass correlation and Pearson’s correlation coefficient. Categorical variables were described using frequencies and percentages, and continuous variables using medians, interquartile ranges, means, and standard deviations.
Results :
Twenty DME and 20 RVO Cirrus SD-OCT B-scan macular cubes with retinal fluid were analyzed. In the DME cohort, the overall fluid grader-grader agreement (GGA) was 91.9%, while the NOA-grader agreement (NGA) was 76.5%. This further stratified into 92.2%, 98.9% GGA and 82.3%, 80.8% NGA for intraretinal and subretinal fluid (IRF, SRF) respectively. RVO overall fluid GGA was 96.6% and 68% NGA, differentiating into 96.4%, 99.8% GGA and 42.2%, 42.9% NGA in the IRF and SRF categories.
Conclusions :
Manual expert graders exhibited high levels of agreement in both the DME and RVO cohorts, generally higher in evaluating SRF. Between the manual graders and the NOA algorithm a satisfactory level of agreement was found, albeit relatively lower in comparison to human graders. Additionally, there was higher machine-human agreement in the DME cohort than the RVO group. The NOA algorithm is helpful in the classification and quantification of retinal fluid may be a useful clinical diagnostic and management tool to augment decision making.
This is a 2021 ARVO Annual Meeting abstract.