Abstract
Purpose :
To train and characterize the effectiveness of a hybrid deep learning system that combines Optical Coherence Tomography Angiography (OCTA), OCT structural data, and foveal OCT b-scans to distinguish between normal eyes, eyes with non-neovascular age-related macular degeneration (AMD), and eyes with neovascular AMD. To also determine retinal pathology features most predictive of neovascular AMD diagnosis.
Methods :
We used 346 retrospective OCTA scans from patients 18 years and older; 97 were diagnosed with no significant vascular pathology (non-AMD), 169 with non-neovascular AMD, and 80 with neovascular AMD with actionable choroidal neovascularization (CNV). For each patient, an OCTA volume and an OCT structural volume were created using deep retinal, avascular, outer retina and choriocapillaris, choriocapillaris, and choroid layers as input for a 3D convolutional neural network (CNN); foveal OCT b-scans were used as input for a 2D CNN. Hybrid CNNs were constructed and trained respectively on: (1) OCTA, (2) OCTA and OCT structure, (3) foveal OCT b-scans, and (4) OCTA, OCT structure, and foveal OCT b-scans combined (Figure [1]). Each CNN’s performance was evaluated via accuracy, precision, recall, and F-1 score. In addition, multinomial logistic regression analysis was conducted to determine the predictive importance of 5 retinal features (adjudicated for presence by OCTA experts) for final AMD diagnosis by experts and by CNNs: (1) intra/sub-retinal fluid (IRF/SRF), (2) scarring, (3) geographic atrophy, (4) CNV, and (5) pigment epithelial detachment.
Results :
The CNN achieving highest test accuracy of 77.8% for this 3-class detection task combined OCTA and OCT structure. Next came the CNN combining OCTA, OCT structure, and foveal OCT b-scans at 75.2% accuracy. The model combining all 3 modalities had slightly higher precision, recall, and F-1 score for the neovascular AMD class. For CNNs, IRF/SRF was an important predictor for neovascular AMD vs. non-AMD eyes; the CNN was able to predict IRF/SRF presence with up to 82.4% accuracy.
Conclusions :
Just as experts rely on both OCTA and OCT structure to diagnose AMD, CNNs also performed best when trained on OCTA and OCT structure combined. IRF/SRF, known to be an important predictor for neovascular AMD diagnosis by experts, was found to be important by CNNs as well and was detected with high accuracy. [1] Thakoor et al., ISBI 2021
This is a 2021 ARVO Annual Meeting abstract.