Abstract
Purpose :
Anti-vascular endothelial growth factor (anti-VEGF) injection is now the first-line therapy for central-involved Diabetic macular edema (DME). However, the response to anti-VEGF treatment is not predictable. Alternative treatments should be offered to the non-responders as soon as possible to gain earlier and better visual improvement. We developed a temporal convolutional network (TCN) model to predict the change of visual acuity (VA) at one year after 3 monthly anti-VEGF injections using optical coherence tomography (OCT) images at baseline, one- and three-month follow-up visits.
Methods :
Serial OCT images from 317 patients who received at least 3 monthly anti-VEGF treatments, had serial OCT images and completed 12-month follow-up at Taipei Veterans General Hospital were retrospectively collected. The treatment effect was defined by the change of VA at 12-month follow-up (final VA) compared with baseline VA. The improved group refers to patients whose final VA improved for over 2 lines at the Snellen chart compared with baseline. The other patients were classified as non-responders. The image dataset was divided into training and testing datasets at a ratio of 5:1. To apply the TCN on image data, a pre-trained model ResNet50 was used to extract the image features, and then fine-tuned by the training dataset. (Figure)
Results :
A total of 101 patients were classified as the improved group. Owing to imbalanced data distribution, data augmentation was performed specifically in the improved group. If 3 OCT images concatenated in the channel dimension were used to predict treatment outcome using ResNet50, the predictive accuracy, specificity and sensitivity were 69.04%, 70.37%, and 68.05% respectively. However, if TCN was applied to catch time-series information of the OCT images, the predictive accuracy improved to 81.25%, with a specificity of 74.40% and sensitivity of 92.07%.
Conclusions :
Applying TCN to extract serial features of OCT images following anti-VEGF treatment may be a workable architecture to predict visual outcomes following anti-VEGF treatment in DME patients. This approach may help us to identify unresponsive patients who may benefit by switching to steroid treatment or surgical intervention to avoid unnecessary continual anti-VEGF injection.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.