Impact of AI architecture on GA area: Comparison of 12 models

Abstract

Purpose : AI algorithms have shown impressive performance in segmenting geographic atrophy (GA) from fundus autofluorescence (FAF) images. Precise identification of hypoautofluorescent boundaries is important for accurate GA segmentation and progression measurement. Selection of AI architecture, and the combination of encoders and decoders is an important variable. In this study, we explore twelve distinct AI architecture combinations to determine the most effective approach for GA segmentation algorithms

Methods : We investigated various AI architectures, each with distinct combinations of encoders and decoders. The model architectures, including U-Net, PSPNet (Pyramid Scene Parsing Network), and FPN (Feature Pyramid Network), serve as foundation framework for segmentation task. The encoder-decoders such as EfficientNet, ResNet (Residual Networks), VGG (Visual Geometry Group) and Mix Vision Transformer (mViT)have a role in extracting optimum latent features for an accurate GA segmentation. These combinations resulted in a diverse set of architectures, totaling twelve models. Model performance was assessed by comparison of area of GA between human ground truth and AI prediction along with Dice Coefficient (DC), in a 5-fold cross validation framework. In addition, masked graders evaluated model performance subjectively using a 4-point scoring system.

Results : The training dataset included 601 FAF images from AREDS2 study and validation included 156 FAF images from GSK BAM study. The mean absolute difference between grader measured and AI predicted areas ranged from -0.08 (95%CI:-1.35,1.19) to 0.73mm (95%CI:-5.75,4.29) and DC ranges between 0.884-0.933. There was no difference of model performance based on GA phenotype, e.g. multi vs unifocal, sub foveal vs extrafoveal. The best performing models with least area difference, DC > 0.9 and grader score of 1 or 2 were UNet and FPN architectures with mViT encoder-decoder.

Conclusions : The choice of AI architecture significantly impacts GA segmentation performance. Vision transformers with FPN and UNet architectures demonstrate stronger suitability for this task compared to CNN and PSPNet based models. The choice of architecture should align with the specific objectives and consider the unique challenges posed by GA analysis.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.