Abstract
Purpose :
To develop a fully automated algorithm based on deep learning to detect and measure atrophy extent in atrophic age-related macular degeneration (AMD) based on macular spectral-domain optical coherence tomography (SD-OCT) volume scans.
Methods :
SD-OCT volumes from 64 eyes (59 patients) with atrophic AMD were extracted and each b-scan was manually annotated for the zone depicting retinal pigment epithelium and outer retinal atrophy (RORA), both for complete and incomplete atrophy. Fortysix SD-OCT volumes were used to develop an automated algorithm based on convolutional neural network to detect and quantify both atrophy types. 18 sets of SD-OCT were kept aside for testing the performance of the algorithm against manual segmentation by two experts.
Results :
The fully-automated algorithm achieved good performance for detection of complete atrophy with a sensitivity of 90%, a specificity of 94%, a precision of 83% and an area under the curve (AUC) of 0.98. The detection of incomplete atrophy was excellent with a sensitivity of 94%, a specificity of 98%, a precision of 94% and an AUC of 0.99. The agreement between each expert and the deep learning model was confirmed with a kappa coefficient of 0.76 (expert 1 - model), 0.79 (expert 2 - model) for complete atrophy and 0.84 (expert 1 - model), 0.89 (expert 2 - model) for incomplete atrophy. Examples of detected atrophy in b-scans are shown in Figure 1 and en face overlays in Figure 2.
Conclusions :
The developed deep learning model achieved performance close to the two human experts for the detection and delineation of atrophy. Detection of complete atrophy proved to be more challenging than incomplete atrophy, which can be explained by the difficulty to establish the exact transition point between those two forms of atrophy. This first use of deep learning to identify complete and incomplete atrophy areas based on SD-OCT scans has the potential of a quick, precise, and highly reproducible OCT-based measurement of atrophy in AMD.
This is a 2020 ARVO Annual Meeting abstract.