Abstract
Purpose :
Age-related macular degeneration (AMD) is projected to affect an average of 1.23 million individuals by the 2050. Whilst anti-VEGF treatment for neovascular AMD (nvAMD) is considered the current gold-standard care, this requires regular monitoring and treatment delivery which causes increased capacity challenges. This, along with the current COVID-19 pandemic, have highlighted the need for efficient and safe ways to diagnose and manage nvAMD. The use of artificial intelligence (AI) in medical care has the potential to alleviate some of this projected pressure facing eye clinics. Previous research has shown that AI has comparable sensitivity and specificity to clinicians in identifying ocular disorders from retinal images. The purpose of the current study was to develop and AI model to identify active from inactive nvAMD disease from retinal SD-OCT images.
Methods :
Using Google’s Vision AutoML software, 1058 Heidelberg SD-OCT images were identified and labelled as either showing nvAMD activity or inactivity. All images were uploaded to Google’s cloud storage and automatically assigned two bounding-box labels; 1 label capturing the entire Heidelberg SD-OCT image, including the raster and b-scan, with the second capturing the b-scan only. All labels were automatically allocated to either a train, validate or test group based on an 80:10:10 ratio set by the software.
Results :
Of the 1058 images, a total of 2116 labels were assigned, 1012 showing active and 1104 showing inactive nvAMD. Performance of the AI model revealed an area under the precision recall curve (AUPRC) of 0.84 at a threshold of 0.5, specificity of 40.98% and sensitivity of 95.24%. For the active-only images, the specificity was 34.28% with a sensitivity of 97%. For the inactive-only images, the specificity was 51% with a sensitivity of 92.73%.
Conclusions :
Utilising Google’s AutoML AI software, this model is able to correctly identify active nvAMD from Heidelberg SD-OCT images with a high level of sensitivity and good overall AUPRC.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.