Deep Learning to Predict the Future Growth of Geographic Atrophy from Fundus Autofluorescence

The region of growth (ROG) of geographic atrophy (GA) throughout the macular area has an impact on visual outcomes. Here, we developed multiple deep learning models to predict the 1-year ROG of GA lesions using fundus autofluorescence (FAF) images.

In this retrospective analysis, 3 types of models were developed using FAF images collected 6 months after baseline to predict the GA lesion area (segmented lesion mask) at 1.5 years, FAF images collected at baseline and 6 months to predict the GA lesion at 1.5 years, and FAF images collected 6 months after baseline to predict the GA lesion at 1 and 1.5 years. The 1-year ROG from the 6-month visit was derived by taking the difference between the GA lesion area (segmented lesion mask) at the 1.5-year and 6-month visits.

Patients enrolled in the following lampalizumab clinical trials and prospective observational studies: NCT02247479, NCT02247531, NCT02479386, and NCT02399072.

Datasets of study eyes from 597 patients were split into model training (310), validation (78), and test sets (209), stratified by baseline or initial lesion area, lesion growth rate, foveal involvement, and focality. Deep learning experiments were performed using the 2-dimensional U-Net; whole-lesion and multiclass models were developed.

The performance of the models was evaluated by calculating the Dice score, coefficient of determination (R²), and the squared Pearson correlation coefficient (r²) between the true and derived GA lesion 1-year ROG.

The model using baseline and 6-month FAF images to predict GA lesion enlargement at 1.5 years had the best performance for the derived 1-year ROG. Mean Dice scores were 0.73, 0.68, and 0.70 in the training, validation, and test sets, respectively. The R² (0.77, 0.53, and 0.79) and r² (0.83, 0.61, and 0.79) showed similar trends across the 3 sets.

These findings show the potential of using baseline and/or 6-month visit FAF images to predict 1-year GA ROG using a deep learning approach. This work could potentially help support decision-making in clinical trials and more informed treatment decisions in clinical practice.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.