Deep learning models predict growth of geographic atrophy lesions using FAF images
Deep learning models using baseline and 6-month fundus autofluorescence (FAF) images can accurately predict the 1-year growth of geographic atrophy (GA) lesions in the macular area, according to a study.
The retrospective analysis included 597 patients from lampalizumab clinical trials and observational studies. Researchers developed 3 types of models:
- Using 6-month FAF images to predict GA lesion area at 1.5 years.
- Using baseline and 6-month FAF images to predict GA lesions at 1.5 years.
- Using 6-month FAF images to predict GA lesion areas at both 1 and 1.5 years.
The dataset was divided into three groups: one for training the models, one for fine-tuning them, and one for testing how well they worked. The models used a type of deep learning method called 2D U-Net. Their accuracy was measured using scores that show how closely the predictions matched the actual outcomes.
The best results came from the model that used images taken at the start and after six months. It predicted 1-year lesion growth with accuracy scores of 0.73, 0.68, and 0.70 for the training, fine-tuning, and testing groups, respectively. Other measures of accuracy also showed strong results, with a final test score of 0.79 for both key metrics.
Refernece
Salvi A, Cluceru J, Gao SS, et al. Deep Learning to Predict the Future Growth of Geographic Atrophy from Fundus Autofluorescence. Ophthalmol Sci. 2024;5(2):100635. doi: 10.1016/j.xops.2024.100635. PMID: 39758130; PMCID: PMC11699103.
This content is independent editorial sponsored by Astellas. Astellas had no input in the development of this content.
Related Content
