Deep Learning based Algorithm for Automatic cRORA and Photoreceptor Loss Detection in SD-OCT Imaging

Abstract

Purpose : Spectral-domain optical coherence tomography (SD-OCT) image features underlying geographic atrophy (GA) progression could provide valuable insights correlating with GA growth measured in FAF. This work aims to develop an automatic deep-learning based algorithm to delineate and quantify regions with complete RPE and outer retina atrophy (cRORA) and photoreceptor loss in (PRL) SD-OCT imaging.

Methods : SD-OCT volumetric scans (512x128 B scans from Zeiss Cirrus) from GA patients were retrospectively acquired from routine clinical care as part of an IRB-approved study at Cleveland Clinic Foundation. Trained graders manually annotated areas with cRORA and PRL following CAM-group guidelines in 3456 B scans from 20 patients (27 eyes). An instance-segmentation based deep learning model was trained to detect and annotate regions with cRORA and PRL. B scan level detections were aggregated to construct an enface map capturing the total area with cRORA and PR loss for each volume scan. Algorithm performance was evaluated in a cross-validated setting by computing precision, recall, Dice score, and area difference. In a separate experiment, intra-reader variability in ground truth was assessed by repeatedly annotating 25 B scans by 4 annotators.

Results : The segmentation algorithm identified areas with cRORA with mean (± stdev) precision = 0.83 ± 0.14, recall = 0.88 ± 0.10, and Dice = 0.84 ± 0.10. PRL segmentation demonstrated comparable performance with precision = 0.85 ± 0.10, recall = 0.87 ± 0.12, and Dice= 0.85 + 0.8. Area difference between algorithm output and human annotations were 1.2 ± 1.9 mm² for cRORA and 1.6 ± 2.0 mm² for PRL. The Dice coefficient among human annotators was found to be 90.1 ± 0.10 for cRORA and 82.3 ± 0.19 for PRL.

Conclusions : Overall, automatic segmentation algorithms performed comparably to human annotations providing a time/cost effective, quantitative method for analyzing SD-OCT images compared to human interpretation. Automatic OCT measurements could be deployed in large data sets to understand the factors underlying GA progression and establish correspondence with GA growth measured in FAF. The reported performance metrics with additional validation could enable successful translation of the algorithm in future clinical studies for automatic computation of GA related metrics.

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