Abstract
Purpose :
The presence of drusen is the hallmark of early/intermediate age-related macular degeneration (AMD), and their sudden regression is strongly associated with the onset of late AMD. In this work, we develop and evaluate a machine learning algorithm to predict individual drusen regression using optical coherence tomography (OCT) based biomarkers.
Methods :
Patients with early/intermediate AMD were followed in a standardized manner every 3 months for a minimum of 15 months. Imaging was performed using Spectralis OCT (scanning area 20°x20°, volume scan 1024x97x496 voxels). Segmentation of retinal layers was obtained using a graph-theoretic approach (Iowa Reference Algorithms). Layer segmentation errors in the outer retina were corrected by an expert reader. In addition, hyperreflective foci (HRF) were segmented by voxel classification. Then a series of automated image analysis steps were applied to identify and characterize individual drusen and their development from baseline to the next follow-up (month 3). The point of regression of each druse was determined algorithmically as a time-point when its volume fell below 10% of the baseline value. Finally, a machine learning algorithm based on random forest was employed to predict the occurrence of drusen regression within the subsequent 12 months (Figure 1).
Results :
The predictive model was trained and evaluated on a longitudinal OCT dataset of 61 eyes from 38 patients using leave-one-patient-out cross-validation. A total of 938 drusen were identified, out of which 71 regressed (7.6%) during follow-up. The prediction performance was quantitatively evaluated as area under the curve (AUC). Detection of individual drusen regression within 12 months achieved an AUC = 0.76, corresponding to a sensitivity of 0.73 and a specificity of 0.70. The presence of HRF was found to be an important biomarker for incoming drusen regression.
Conclusions :
The predictive model proposed in this pilot study represents a promising step toward image-guided prediction of AMD progression. Machine learning will substantially contribute to the development of new therapeutics that target slowing the progression of early/intermediate AMD.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.