Diagnostic accuracy of a machine-learning algorithm to detect and classify choroidal neovascularization based on SD-OCT in neovascular age-related macular degeneration (nAMD)

Abstract

Purpose : To evaluate the performance and accuracy of a machine learning (ML) algorithm to detect and classify choroidal neovascularization (CNV) based on SD-OCT images when compared to fluorescein angiography (FA).

Methods : Baseline FA and SD-OCT B-Scans from study and fellow eyes in the HARBOR study (NCT00891735) were used to develop, train and cross-validate a ML pipeline combining deep learning for segmentation and classification based on features. B-Scans from the AVENUE study (NCT02484690) were used to externally validate an ML model built on all images from HARBOR. FA images were only used to provide the ground truth of the diagnosis for the model development.

Results : Images of 1037 study eyes of 1098 patients with CNV, as well as 531 images of fellow eyes without CNV from the HARBOR study were used to train and test multiple models in a five-fold crossvalidation setting. The algorithm discriminated absence of CNV from presence of CNV with almost perfect accuracy, AUROC=0.99 (95% CI 0.99 – 1.00), and CNV types “Occult” and “Classic” as per FA assessment with high accuracy AUROC=0.91 (95% CI 0.89-0.94) on the HARBOR SD-OCT images. CNV types could be discriminated with AUROC=0.88 (95% CI 0.82-0.95) on baseline SD-OCT images of 165 study eyes with CNV from AVENUE.

Conclusions : We developed and evaluated an ML model to detect presence and type of CNV in patients with treatment naïve nAMD on SD-OCT images with high accuracy. The results of this pilot study are an important step forward towards automated noninvasive identification and classification of CNV and provide an effective tool for monitoring the evolution of CNV in both clinical trials and tele-ophthalmology programs

This is a 2020 ARVO Annual Meeting abstract.

 

Diagnostic accuracy of a machine-learning algorithm to detect and classify choroidal neovascularization based on SD-OCT in neovascular age-related macular degeneration (nAMD)