Feature discovery using ablation studies for deep learning–based geographic atrophy (GA) progression prediction

Abstract

Purpose : Although convolutional neural networks (CNN) can accurately predict GA growth rate using fundus autofluorescence images (FAF), what drives the prediction accuracy remains unknown. Here, we ablate different regions of FAF to study their contribution to GA progression prediction.

Methods : The dataset included baseline (BL) FAF of patients with bilateral GA enrolled in NCT02479386, NCT02247479, and NCT02247531. GA progression, measured by annualized lesion growth rate (mm²/y), was derived as the slope of a linear fit on available measurements of GA lesion area. Data were split into training (n=1041) and holdout (n=255) sets stratified by BL factors. The training set was further split into 5 stratified folds.
GA lesion masks were generated using a segmentation algorithm for each image. Three regions were delineated using the mask: inside the lesion (Lesion), a 500-µm rim around the lesion (Rim), and the remaining region. All possible combinations of these regions were ablated with black pixels to create new datasets. The VGG16 architecture was trained using a random hyperparameter search on each dataset to predict GA progression. The square of the Pearson correlation coefficient (r²) between predicted and observed GA growth rate was averaged across the 5 folds and used for model selection. The same metric was then used to test performance on the holdout set.

Results : Each 5-fold cross-validation experiment yielded 5 independent models. Fig 1 depicts the holdout r² of each of these models for each dataset.
Compared with the full FAF (r²=0.44), nearly half of the variability in prediction was explained by capturing only shape and size in the Mask Only dataset (r²=0.24). The No Rim dataset (r²=0.39) showed a greater decrease in performance compared with the Lesion and Rim (r²=0.43) or No Lesion (r²=0.42) datasets. Further, the Rim Only dataset (r²=0.37) performed better compared with the Lesion Only (r²=0.33) and No Lesion or Rim (r²=0.27) datasets.

Conclusions : The intensity and texture within the rim of the FAF contains more predictive features for GA growth rate. We interpret that intensity/texture and morphology have additive contributions to the predictive performance of the CNN. Our findings can be further confirmed by quantifying these features using oculomics and observing predictive performance.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Fig 1. GA progression is shown with representative images of each ablation technique