Comparison of a Deep Learning based OCT image segmentation algorithm to manual segmentation by a traditional reading center for patients with wet AMD

Abstract

Purpose : Imaging features determined by human graders are currently the gold standard in retinal imaging. In this study we assessed the accuracy of artificial intelligence (AI)-based tools in quantifying retinal biomarkers from Optical Coherence Tomography (OCT) images as a time efficient alternative to a human.

Methods : This is a retrospective study of patients (n = 250) with wet age-related macular degeneration (AMD) derived from the American Academy of Ophthalmology’s IRIS registry. Two OCT volume scans from the study eye were obtained in each patient: at baseline and at the end of 2-year follow-up. Central retinal thickness (CRT) was measured by a human grader, and central subfield thickness (CST) was measured by fully autonomous deep learning (DL) algorithm (RetinAI Medical AG, Switzerland). Subretinal fluid (SRF) and intraretinal fluid (IRF) volume were estimated both by a human grader and DL algorithm. To assess correlation and agreement, Spearman correlation (r₂) and Bland-Altman analyses were performed (Fig. 1 A-C).

Results : SRF and IRF were present and estimated in 180 (36.0%) and 139 (27.8%) OCT volumes, respectively (Fig 1. D-F). A significant correlation between the volume estimates by a human grader and the DL algorithm was observed. This held true for both SRF (r₂: median = 0.96; 95% CI = 0.93-0.98) and IRF (r₂: median = 0.78; 95% CI = 0.67-0.86), but IRF was slightly more variable. The CST estimates by the algorithm were significantly higher, as expected due to the convex shape of the region, but the two were significantly correlated to each other (r₂: median = 0.89; 95% CI = 0.85-0.93) (Table 1).

Conclusions : There was strong agreement between the DL-based algorithm and the gold standard (a human grader) in this wet AMD cohort. This study provides evidence that DL-based tools can help improve trial efficiency by allowing for faster read-outs of study endpoints and potentially help generate faster insights for patient surveillance in trials and real-world.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Figure 1: Bland-Altman plots and b-scans. Points in A-C represent optical coherence tomography (OCT) volumes. The red and green lines in A-C correspond to the median difference and the 25th/75th percentiles of the difference. Areas of pathology in E-F are highlighted in green if detected by both grader and DL algorithm, and red if detected by the algorithm only.

 

Table 1