Abstract
Purpose :
To explore the contribution of clinical and derived fundus autofluorescence (FAF) imaging features in the models for predicting geographic atrophy (GA) growth rate.
Methods :
This study used FAF images (study eye, baseline visit, all treatment arms) from 1281 patients with bilateral GA enrolled in lampalizumab phase 3 trials (NCT02247479/NCT02247531), and an observational study (NCT02479386). Only the macula centered 30° FAF image (Heidelberg Engineering) was used. GA lesion growth rate was computed from the slope of a linear fit on all available measurements of GA lesion area (from FAF images graded by multiple readers). The GA lesion area was obtained using a previously described GA lesion segmentation algorithm.1 Multiple shape features were derived from the segmentation lesion mask. For multifocal lesions, circularity and Feret diameters were quantified using the average of all individual lesions weighted by area. Table 1 lists feature sets derived from the clinical features obtained at study baseline (Study) and from the mask (Shape); when combined, redundant features were removed (Study+Shape). A nested cross-validation (CV) approach was used to tune the model hyperparameters using the inner CV folds for each of the 5 outer folds; 5 modeling strategies (linear models [Lms], Lms with regularization [ElasticNet, Lasso, Ridge], and XGBoost [XGB]) were used on the feature sets. The mean R2 values on the outer folds were compared.
Results :
Fig. 1 shows the mean R2 on the outer folds in the CV. Lm with no regularization was the highest-performing algorithm for all feature sets, followed by XGB. Across all modeling strategies, the Study +Shape feature set performed the best. Shape alone performed the worst, followed by the Study data alone. The best model overall was the Study+Shape Lm (mean R2 = 0.29), followed by the Study+Shape XGB.
Conclusions :
GA lesion shape and clinical features together achieved greater performance than either feature type alone. This suggests use of complementary information provided by imaging and clinical features to predict GA progression. Further work is needed to robustly identify the most relevant shape and clinical features, and to explore other imaging features such as intensity and texture.
Reference
1. Spaide et al. Invest Ophthalmol Vis Sci. 2021;62(8):2124.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.