Comparative analysis of training approaches for deep learning geographic atrophy segmentation models

Abstract

Purpose : Deep learning (DL) models have been used for geographic atrophy (GA) segmentation in optical coherence tomography angiography (OCTA) en face images. To understand the impact of variations in DL models and training datasets on GA segmentation, we compared two DL models trained on spectral domain (SD) and swept-source (SS) OCT/OCTA images.

Methods : The DL model from University of Washington (UW)[1] was trained on SS-OCTA en face images from 153 GA eyes and 90 non-GA eyes using both 6x6 mm scans and 12x12 mm scans obtained on PLEX® Elite 9000 SS-OCT (ZEISS, Dublin, CA), and SD-OCTA scans of 40 GA eyes with a 6x6 mm scan pattern on CIRRUS® 5000 (ZEISS, Dublin, CA). The DL model from ZEISS [2] was trained on SS-OCTA en face images from 173 GA eyes and 21 non-GA eyes with both 6x6 mm scans and 12x12 mm scans obtained on PLEX Elite 9000 and transfer learning [3] was applied to a dataset of SD-OCT scans from 150 GA eyes and 24 non-GA eyes with Macular cube 512x128 and 200x200 CIRRUS 5000 scans. 20 GA eyes with SD-OCTA 6x6 mm scan and SS-OCTA 6x6 mm scans formed the testing dataset for model performance evaluation. All GA lesions were manually annotated (MA) from the subretinal pigment epithelium using the same method for both DL models. GA area was calculated for MA and both DL models. Similarities of MA and DL outputs were compared with the Dice similarity coefficient (DSC). An ANOVA test was run to compare mean areas and a paired t-test was used to compare DSC scores.

Results : No significant difference was found in GA area between MA and both DL models for either SD or SS-OCTA images (p > 0.05). In SD-OCTA images, the DSC (+/- SD) between MA and the UW model was 0.88 ± 0.08 and between MA and the ZEISS model was 0.83 ± 0.16. In SS-OCTA images, the DSC between MA and the UW model was 0.77 ± 0.32 and between the MA and ZEISS model was 0.88 ± 0.12. There was no significant difference between the DSC of the two DL models for either SD or SS-OCTA images (p >0.05).

Conclusions : DL models with varying proportions of SD-OCT and SS-OCT images in the training dataset and different model architectures can both successfully segment GA lesions, showing the robustness of DL models trained for GA segmentation.

References:
1. Chu Z et al. BioOptExp 2022. PMID: 35414972.
2. Pramil V et al. Retina 2023. PMID: 35970318.
3. Zhang Q et al. IOVS 2023. 64(8):1109.

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