Abstract
PURPOSE: To compare geographic atrophy (GA) size measured with fundus autofluorescence (FAF), near-infrared (N-IR) imaging, retromode (RM) imaging and optical coherence tomography angiography (OCTA) imaging and to compare accuracy of artificial intelligence(AI)-based automatic segmentation of GA with each method. METHODS: Available good quality FAF, N-IR- RM and OCTA images acquired on the same date for each patient diagnosed with GA from 2022 to 2024 were retrospectively collected. Seventy(70)% of the images were used to train a Trainable Weka Segmenter (v 3.3.2) based on manual segmentation of GA and spurious areas performed by 2 different blinded expert graders for each of the 4 imaging modalities. For the remaining 30%(testing set), automatic measurement and manual measurement were compared to determine accuracy of the segmentation. RESULTS: A total of 157 eyes were included. Mean ground truth GA area (graders' manual contouring), mean automatic area and mean spurious area of testing set were significantly different with the 4 techniques(respectively p < 0.001, p < 0.001 and p = 0.002). Intraclass correlation coefficient(ICC) between manual and automatic measurements was 0.82 (0.78-0.84) for FAF model, 0.81 (0.78-0.82) for N-IR model, 0.67 (0.64-0.71) for RM model and 0.77 (0.73-0.81) for OCTA model. CONCLUSION: We report very good performance of automatic segmentation performed on FAF, N-IR and OCTA. A slight overestimation of GA area with automatic measurements would be considered when assessing GA area on FAF and N-IR imaging. RM imaging should not be considered as a valid method for automatic GA area assessment due to superiority of other available enface imaging techniques.