Purchase this article with an account.
Ayesha Nuri Karamat, Giulia Corradetti, Nadav Rakocz, Jeffery N Chiang, Muneeswar Gupta nittala, David S Boyer, David Sarraf, Eran Halperin, Srinivas R Sadda; Prediction of Activity in Eyes with Macular Neovascularization Due to Age-related Macular Degeneration Using Deep Learning. Invest. Ophthalmol. Vis. Sci. 2022;63(7):206 – F0053.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Optical coherence tomography angiography (OCTA) can capture the microvascular structure of macular neovascularization (MNV) in eyes with age-related macular degeneration (AMD).In this study, we evaluated several deep learning algorithms designed to detect activity of MNV using en face OCTA images.
En face OCTA 6x6 mm images from 97 subjects with neovascular AMD imaged on RTvue-XR Avanti SD-OCTA (Optovue. Inc, Fremont, CA) device were included in this study and retrospectively analyzed. En face OCTA images of the MNV lesion were generated using a customized 10 microns thick slab,with the boundaries adjusted to display the maximum extent of the MNV lesion. Multiple machine learning models were trained to classify the presence of MNV activity on the OCTA images, using the presence of fluid on the structural OCT as the ground truth. Specifically, a five-fold cross-validation was applied to assess the different models’ performance:four-fifths of the patients were used for training and a fifth for validation. This process was computed 20 times to generate a total of 100 different receiver operating characteristics (ROCs).The different performances were evaluated by using the ROC,its area under the curve (AUC). To further assess the ability of the algorithms to detect activity of MNV using en face OCTA images only, a leaky cross-validation approach was used.
For 97 patients participated 637 en face OCTA scans were exported as .png images. Macular neovascularization (MNV) evident on en face OCTA images was a poor predictor of disease activity as defined by the presence of fluid on structural OCT. The algorithms used did not have a good performance and using the leaky cross-validation approach we conclude that even if a larger number of cases were used a substantial improvement in performance is not expected: Resnet (0.59 [0.47,0.71]), simple CNN (0.60[0.49,071]), LR+PCA (0.53[0.41,0.64]), Resnet-Scratch (0.62[0.54,0.70]). The leaky cross-validation applied to the top performing model, Resnet-Scratch, resulted in an AUROC of 0.67[0.60, 0.74].
In this study, we observed that en face OCTA images alone of MNV lesions are poor predictors of MNV lesion activity.This suggests that strong biomarkers of disease activity at a point time may not be encoded within the en face OCTA image.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
This PDF is available to Subscribers Only