Post-treatment Visual Acuity Prediction Using Deep Learning in Age-related Macular Degeneration
10.3341/jkos.2023.64.7.582
- Author:
Najung KIM
1
;
Hyung Chan KIM
;
Hyewon CHUNG
;
Hyungwoo LEE
Author Information
1. Department of Ophthalmology, Konkuk University Medical Center, Konkuk University School of Medicine, Seoul, Korea
- Publication Type:Original Article
- From:Journal of the Korean Ophthalmological Society
2023;64(7):582-590
- CountryRepublic of Korea
- Language:Korean
-
Abstract:
Purpose:To develop a deep learning model to predict visual acuity (VA) outcomes after 12 months of anti-vascular endothelial growth factor (anti-VEGF) treatment.
Methods:A total of 330 treatment-naive eyes of neovascular age-related macular degeneration patients, who underwent anti-VEGF therapy between 2007 and 2020 at Konkuk University medical center, were included. The network was trained using VA at baseline, VA after three loading doses of anti-VEGF, and treatment regimen data. It was also trained using 12,300 augmented optical coherence tomography (OCT) B-scan images at baseline and after three loading doses of anti-VEGF. We generated five deep learning models using sequentially input data (VA and OCT B-scan images at baseline and after three loading doses, and treatment regimen). Prediction of VA at 12 months was performed using deep learning algorithms, such as convolutional neural network and multilayer perceptron. The outcomes were dichotomized based on whether the decremental change in VA during the 12 months of treatment was more or less than logarithm of the minimum angle of resolution 0.3. Predictive efficiency was assessed by comparing the performance of deep learning models.
Results:The best performing model was trained using input data, including VA at baseline and after three loading doses, treatment regimen, and OCT B-scan images at baseline and after three loading doses. The decremental outcome in VA after 12 months of anti-VEGF treatment was predicted as an area under the curve (AUC) of 0.79. The addition of OCT images at baseline and after three loading doses as input data improved the AUC, sensitivity, and negative predictive value (AUC 0.74-0.79, 0.58-0.86, and 0.90-0.95, respectively).
Conclusions:Our deep learning model showed relatively good performance in classifying good or poor post-treatment VA based on combined clinical information including numerical and image data.
