Abstract
Purpose :
To demonstrate a deep learning model for automatic segmentation of geographic atrophy (GA) using en face OCT projection images from various OCT scan patterns and different OCT instruments: 840 nm spectral domain OCT (SD-OCT) and 1060 nm swept-source OCT (SS-OCT).
Methods :
Pseudo-color images characterizing the main OCT features of GA were generated from the volume data using an automated algorithm that includes hyper-transmission in sub-retinal pigment epithelium (RPE) slab, regions of RPE loss, and loss of retinal thickness as training inputs. A 5-fold cross-validation model with data augmentation techniques was trained with 126 Angio 6x6 mm scan images from PLEX® Elite 9000 SS-OCT (ZEISS, Dublin, CA), 80 Angio 12x12 mm PLEX Elite images, and 188 Macular Cube 512x128 or 200x200 scan images from CIRRUSTM HD-OCT 5000 SD-OCT (ZEISS, Dublin, CA) (Figure 1). GA lesions were manually outlined by expert graders as ground truth for training and evaluation. Sensitivity, specificity, and two-tail dice coefficient were calculated and used as the evaluation metrics of the trained model. GA area and the square root area were measured. A test set of 10 images for each scan pattern from 10 eyes with GA as well as 10 images from 10 eyes without GA were tested and evaluated.
Results :
The output of the model was obtained by simply averaging the outputs from the 5 trained models with optimal thresholds. A high sensitivity of 0.88, a high specificity of 0.99 and a high two-tail dice coefficient of 0.93 in the training set and a sensitivity of 0.95, a specificity of 0.85 and the two-tail dice coefficient of 0.91 in the validation set were achieved. There were no significant differences in GA area between the automatic GA area measured from the model and the manual outlines from the human graders across all the scan patterns and instruments.
Conclusions :
The single deep learning model that incorporates various scan patterns and different OCT instruments was able to accurately delineate GA lesions with a high sensitivity and specificity. This model could serve as a new management tool for GA detection across multiple scan patterns and OCT devices.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.