Abstract
Purpose :
To develop a machine learning(ML) aided pipeline to automatically confirm geographic atrophy(GA) using real-world fundus autofluorescence(FAF) and infrared reflectance(IR) images linked to the IRIS Registry
Methods :
The American Academy of Ophthalmology IRIS® Registry(Intelligent Research in Sight)is the nation’s first comprehensive eye disease clinical database, with over 70% of US ophthalmologists contributing. A ML aided pipeline was developed for the IRIS Registry, using a de-identified clinicoimaging dataset with images contributed by 2 large retina practices, patients with a dry age-related macular degeneration(AMD) diagnosis were identified from 2006 to 2019. ML models were developed to assess FAF and IR image quality using features of signal, contrast, noise and sharpness. FAF images with a predicted quality score lower than 0.4[range: 0-1] were excluded. For patients with multiple images at the same encounters, or multiple encounters, the image with the highest quality score was selected. Lastly, training and testing sets were selected using patient-level stratified sampling(age, sex, and race) and labeled by a trained grader. Two deep learning models were developed to classify patients into three categories: no GA, GA without subfoveal involvement, and GA with subfoveal involvement. The models were trained on the training sets with an 80:20 split for training and validation. The trained models were then deployed on the entire cohort
Results :
In total, 15,023 FAF and 16,470 IR images from 4,372 eyes of 2,248 patients were identified. After removing low quality images(1,931 FAF images) and multiple images from the same visit(6,825 FAF images, 8,285 IR images), 6,267 pairs of FAF and IR images from 3,372 eyes of 1,872 patients were included. A training set of 442 patients and a testing set of patients(one pair of FAF and IR images per patient) were selected from the cohort. A modified VGG network was developed for GA diagnosis. The model achieved an accuracy of 0.96, 0.96 and 0.94 for training, validation and testing, respectively. The second model for subfoveal involvement is under development
Conclusions :
The proposed pipeline demonstrated satisfactory performance confirming GA in AMD eyes using images collected in routine practice. It could potentially be useful in screening patients for GA trials and future GA treatments
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.