Abstract
Purpose :
Late AMD, either dry or exudative, is a leading cause of irreversible blindness. Preventing progression of early stage AMD is thus very important for disease management. We focus on the development and validation of an AMD prediction model based on color fundus photography and an individual’s socio-demographic parameters to provide a risk score for progression to late AMD.
Methods :
We used 7611 images (total 3535 cases) from the AREDS (Age-Related Eye Disease Study) dataset for the development and validation of the model. Out of these, 2381 images (1188 eyes) belong to 923 individuals who had either one or both their eyes progress to late AMD during the study, the images chosen from the visit prior to progression. The rest of the images (total 5230) belong to 2665 individuals (normal or early AMD) who never suffered from advanced AMD during this study. Our proposed model is based on pathology segmentation and AMD severity level computation from retinal color photography. We combined these parameters with demographic factors such as age, race, sex, presence of diabetes, BMI and sunlight exposure, i.e., with traditional AMD risk factors. Initially, the color channel transformation from RGB to CIE L*a*b is applied to generate perceptually uniform color space . We applied standard edge based image segmentation (canny edges, graph based region growing of edge pixels, etc) to identify and quantify the pathologies. Following this, deep convolution neural network models were applied to generate a risk score from 0 to 1 of progressing to late AMD within 6-months to two years. Finally, a decision tree model was applied to the fusion of the neural networks results and pathology quantifications to finalize the risk score; a score above 0.5 predicting progression. We split the dataset into roughly 80% for training and 20% for testing.
Results :
We achieved a sensitivity of 89% and a specificity of 77.4% in predicting individuals who would progress to late AMD in 6 months to 2 years. We could correctly identify 487 out of the 546 images from the individuals who developed late AMD. For individuals who never converted, it could correctly identify 825 out of 1066.
Conclusions :
We have developed a model for predicting individuals who will develop late AMD in the near future, i.e., six months to 2 years. We aim to make this tool available to ophthalmologists for the best care in the prevention of incident late AMD.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.