Abstract
Purpose :
Using deep learning (DL), this pilot study aimed to demonstrate that function in the form of visual acuity can be predicted from pathologic structural alterations due to neovascular age-related macular degeneration (nAMD) in optical coherence tomography (OCT) images.
Methods :
This retrospective analysis included 2443 OCT volumes from 341 eyes of 294 patients with nAMD. Only the foveal line scan from each OCT volume was used. We train a deep convolutional neural network using transfer learning for the binary task of predicting better or worse than 20/40 BCVA. A cut off VA of ≥ 20/40 was chosen, as it is the accepted cut-off value for driving in most states in the US and ≥ 20/40 has been used in several large clinical trials to characterize good visual outcomes in nAMD patients. The BCVA recorded at each corresponding clinic visit was used as ground truth. The entire dataset (BCVA ≥ 20/40 1097 images; BCVA < 20/40 1346 images) was split randomly at the patient level into training (82.8%), validation (12.3%) and testing (4.9%).
Results :
For the binary classification task of distinguishing between better and worse than BCVA 20/40, our DL algorithm achieved an accuracy of 80.0%, precision of 80.0%, recall of 79.0%, with the area under the ROC of 86.2%.
Conclusions :
We have developed a DL algorithm that is capable of predicting function (BCVA) from structure (OCT images) in nAMD. As the next step, we plan to collect more training data and refine our algorithm for more fine-grained BCVA predictions, which if successful, could lead to the development of a novel DL-based OCT metric that can be used to monitor treatment efficacy in nAMD.
This is a 2021 ARVO Annual Meeting abstract.