Validation of SD-OCT Derived Automated Machine Learning-Augmented Volumetric Fluid Quantification

Abstract

Purpose : Accurate detection and quantification of fluid in SD-OCT images can assist with treatment decisions and prognostic evaluations in exudative retinal diseases. This study evaluated the performance of a fully-automated, machine learning (ML)-augmented OCT segmentation platform in quantifying macular fluid volumes in wet age-related macular degeneration (AMD) and diabetic macular edema (DME).

Methods : A total of 120 volumetric SD-OCT scans (60 wet AMD and 60 DME) from both Cirrus and Spectralis devices (1:1 ratio for both diseases) were included. A previously developed ML segmentation platform provided fully-automated segmentation of intraretinal fluid (IRF) and subretinal fluid (SRF). Two trained readers (R1 & R2) independently reviewed and corrected the baseline ML segmentation (semi-automated segmentation) while two senior readers collaborated to create a gold-standard (GS) semi-automated segmentation for reference. Central macular fluid volumes (i.e., central 2 mm foveal-centered zone) from ML, R1, and R2 were compared to GS volumes. Agreement was assessed using intraclass correlation coefficients (ICC) and Bland-Altman plots.

Results : In DME, ML achieved an ICC of 0.976 for IRF compared to reader values of 0.992 for both R1 and R2. In wet AMD, ML achieved an ICC of 0.895 for IRF compared to values of 0.971 and 0.970 for R1 and R2, respectively. For SRF volumes in wet AMD, ML achieved an ICC of 0.988 compared to reader values of 0.997 for both R1 and R2. Bland-Altman plots with mean volume differences and 95% limits of agreement are shown in figure 1.

Conclusions : The ML automated platform demonstrated excellent agreement with the gold-standard achieving performance comparable to the readers in IRF quantification in DME and SRF quantification in wet AMD validating its potential use in future disease characterization. The fully-automated segmentation platform allows for quick and accurate measurements of fluid volumes in SD-OCT scans of DME and wet AMD from two widely used OCT devices.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1: Bland-Altman plots visualizing volume differences for ML and the two readers. Dotted lines indicate mean volume differences and 95% limits of agreement. Differences were calculated by subtracting ML and reader volumes from GS volumes.