Comparative Assessment of a Machine Learning-Enhanced Layer Segmentation Platform to Conventional Image Processing Segmentation for Spectral Domain OCT in Retinal Diseases.

Abstract

Purpose : Optimizing efficiency and accuracy of layer segmentation is important for rapid disease assessment and improving disease understanding. This study examined the comparative performance of a 1^st generation Enhanced Deep Neural Network Algorithm (EDNA) in quantifying retinal layers in an automated fashion in eyes with age-related macular degeneration (AMD), Plaquenil exposure (PLAQ), diabetic macular edema (DME), and eyes without retinal disease (NL) to expert reader (ER)-corrected segmentation and a logic based algorithm (LBA).

Methods : In this image analysis study, 80 NS eyes, 40 AMD eyes, 80 PLAQ eyes, and 77 DME eyes were included. The gold standard metrics for this analysis was based on ER-corrected segmentation scans originally assessed with the LBA. EDNA was developed as an enhancement to the previously described LBA to improve accuracy and efficiency. For this analysis, the inner limiting membrane (ILM) and retinal pigment epithelium (RPE) were segmented. The intraclass correlation coefficients (ICC) for EDNA and the LBA were calculated by comparing their respective segmentation metrics for central subfield thickness (CST) to the ER metrics, and categorized using Cicchetti’s criteria (<0.4 = poor, 0.4 – 0.59 = fair, 0.6 – 0.74 = good, 0.75 – 1.00 = excellent).

Results : EDNA demonstrated good to excellent correlation performance when compared to ER in all disease categories for assessment of retinal thickness (i.e., ILM to RPE). EDNA performance demonstrated higher correlation with ER than the LBA in AMD and PLAQ (ICC CST: 0.60 vs 0.45 ; 0.88 vs 0.61). EDNA also demonstrated similar correlation compared to the LBA when measuring mean retinal thickness in NL (ICC CST: 0.88 vs 0.97). In eyes with DME, EDNA demonstrated a slightly lower correlation to ER than LBA (ICC CST: 0.65 vs 0.81).

Conclusions : EDNA demonstrated good-to-excellent correlation to ER-corrected segmentation across multiple retinal diseases. The next generation EDNA system is currently under development and will be explored for enhanced multi-layer performance while minimizing the requirement for human reader intervention.

This is a 2020 ARVO Annual Meeting abstract.