Abstract
Purpose :
Geographic atrophy (GA) is a late-stage form of age-related macular degeneration (AMD) that is characterized by the loss of retinal layers, including photoreceptors, the retinal pigment epithelium (RPE), and the choriocapillaris (CC). Now, with a recently approved drug for GA, tools for detecting the disease may aid disease management.
Methods :
The purpose of this study was to identify the presence or absence of retinal pigment epithelial and outer retinal atrophy (RORA) 132 atrophic AMD patients from the Medical Retina Department at Jules-Gonin Eye Hospital. The scans were manually segmented for RORA on a b-scan level.
We used an automated 12-layer segmentation algorithm based on b-scans. Layer thicknesses were then computed on an a-scan level. These thickness features were used to train an XGBoost model to classify individual a-scans with the presence of RORA. A total of 113 volumes were used, resulting in 2,926,336 a-scans. The model was trained using 80% of the data, 20% for validation to optimize hyperparameters. The final model was tested on a hold-out set of 19 OCT. No patients were present in both training and test sets.
The model was evaluated using AUC, precision, and recall of identifying RORA, and Dice score for en-face segmentation agreement. SHAP values were used to understand the importance of each feature in predicting RORA. A Wilcoxon test was used to compare layer thicknesses in each a-scan between those with and without RORA.
Results :
The model was able to identify RORA with 92% accuracy, 0.91 AUC and precision and recall of 0.87 and 0.85. An average dice score of 0.78 was achieved in test data.
High SHAP values showed ellipsoid zone, photoreceptors and interdigitation zone (EZ, OPR, IZ); myoid zone (MZ); and choroid (CC, CS), all had high model importance. We also observed that a thinner layer thickness of EZ, OPR, IZ and MZ is highly correlated to the presence of RORA in an a-scan (pvalue: <0.01) whilst conversely, thicker PED and RNFL are correlated with RORA absence (pvalue: <0.01).
Conclusions :
Our research has shown that it is possible to accurately detect the presence of RORA using the thickness of segmented retinal layers and fibrotic regions, without the need for any additional clinical markers. This innovative approach has the potential to improve the quantification of RORA.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.