Abstract
Purpose :
To develop a machine learning (ML)-aided pipeline to identify eyes with geographic atrophy (GA) that have subfoveal involvement and eyes without subfoveal involvement using real-world optical coherence tomography (OCT) images linked to the American Academy of Ophthalmology IRIS® Registry (Intelligent Research in Sight).
Methods :
The IRIS Registry is the nation’s first comprehensive eye clinical database, with over 70% of US ophthalmologists contributing. A ML-aided pipeline was developed using a de-identified clinical-imaging dataset of patients with dry age-related macular degeneration (AMD) identified through ICD-9/10 codes from 2006 to 2022. Images were contributed by 3 different practices. Previously developed ML models that identify eyes with GA secondary to AMD were used to identify eyes with GA. From eyes with GA, a training and testing set were selected using patient-level stratified sampling (age, sex, and race) and labeled by two fellowship-trained retinal specialists. A three round consensus workflow was implemented to ensure both graders agreed on all labels, with the center point of the fovea being involved as the definition of subfoveal involvement. A deep learning model was developed to classify GA with subfoveal involvement from GA without subfoveal involvement, trained on the training set with an 80:20 split between training and validation.
Results :
In total, 265 OCT volumetric images from 256 eyes of 232 patients were labeled. A training set of 178 images, a validation set of 45 images and a testing set of 42 images were randomly split (stratified by subfoveal involvement status) where no patients were included in more than one set. A modified VGG19 network was developed.The model achieved an accuracy of 87%, 84% and 79% for training, validation and testing, respectively. In the testing set, the model achieved a precision of 0.79, a recall of 0.79 and a F-1 score of 0.78.
Conclusions :
The proposed pipeline demonstrated satisfactory performance identifying GA eyes with and without subfoveal involvement using images collected in real world practice. It could potentially be useful in screening patients for GA trials and identifying patients for future GA treatments.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.