Abstract
Purpose :
To address the increasing prevalence of diabetes, artificial intelligence (AI) systems have been developed for diabetic retinopathy (DR) screening and diagnosis. However, current AI systems are limited in predicting the progression and prognosis of DR. Identifying patients more likely to progress and appropriately referring them to an ophthalmologist is key to vision preservation. Our goal was to develop and evaluate a prognostication system that uses longitudinal fundus photos to train deep learning models for predicting DR progression from non-referable to referable status.
Methods :
We conducted a retrospective study comparing pairs of retinal fundus images of the same eye taken at two time points separated by a year or more apart. The fundus photos were obtained from EyePACS DR’s graded dataset. We classified normal, mild nonproliferative DR (NPDR), and moderate NPDR to be non-referable DR (NRDR); severe NPDR and proliferative PDR were classified as referrable DR (RDR). In developing a DR prognostication system, we only considered cases where the first encounter was NRDR. Cases where both encounters were NRDR were labelled as negative; cases progressing from NRDR to RDR were labelled as positive. Using a subset of 12768 images of 6384 eyes (two images per eye, taken 733 +/-353 days apart), we trained a neural network, ResNeXt (80%-10%-10% training-validation-testing split), to predict DR severity. An independent e-ophtha dataset of 148 images with microaneurysm (MA) annotations (75%-25% training-validation split) was used to train another neural network, Mask-RCNN, to estimate the number of existing MAs. The resultant DR severity and MA number scores of the first NRDR image of a given eye were used to predict if the DR had progressed to RDR in the second, later image (Fig 1.)
Results :
The area under the receiver operating characteristic curve (AUC) for predicting RDR was 0.956 with the ResNeXt model. The recall, precision, and F1-score in detecting individual MAs were 0.786, 0.615, and 0.690 respectively with the Mask-RCNN model. A combination of the ResNeXt and Mask-RCNN models had an AUC of 0.989.
Conclusions :
We developed and validated an AI system that prognosticates DR progression based on longitudinal fundus photos. This system has the potential to support clinicians in identifying DR cases more likely to progress and referring them to an ophthalmologist for management.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.