Abstract
Purpose :
Despite enrichment of fast progressors in recent therapy trials based on risk factors, high variability of lesion growth rate between patients with geographic atrophy (GA) remains. We proposed here to apply a deep learning approach on retinal images from internal studies in order to predict the lesion growth rate at 12months from baseline and implement this prediction in future clinical trials.
Methods :
Data were pooled from several internal GA studies representing about 236,822 total images over a period of 1-4years. Images with corresponding lesion size measurements and with follow-up visits were selected. Convolutional Neural Networks (CNNs) were fine-tuned on the datasets and performance of the CNNs were then evaluated using Pearson’s correlation coefficient. We also implemented a modified sensitivity analysis approach to generate activation maps, which allows us to visualize at the pixel-level regions that play a role in the prediction of GA growth.
Results :
The integrated algorithms showed a Pearson correlation coefficient of 0.59. The test dataset was then separated into slow and fast progressors based on the average growth rate for all the pooled trials (0.13mm2/month). This logistic regression approach reached an accuracy of 76.7% and a positive/negative predictive rate of 65.9% and 84.9%, respectively. The activation maps showed that the prediction is mainly based on the lesion itself. Evaluation of the t-distributed stochastic neighbor embedding (t-SNE) map on the training set further suggested that the shape complexity may be the most important risk factor for growth rate.
Conclusions :
We showed here for the first time that baseline retinal images indeed contain predictive information about the GA lesion growth rate at follow-up visits. The high negative predictive value indicate the possibility of screening out slow progressors while the modest positive predictive value suggest that additional parameters may be needed to improve the prediction of fast progressors. Furthermore, the visualization features enabled the validation of the algorithm as well as provided new insights into the natural progression of GA.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.