Calibration of OCT Measurements across Multiple Device Generations Using a Measurement Error Structural Equation Model

Abstract

 

Purpose

 

Retinal nerve fiber layer (RNFL) thickness measurements made by different generations of OCT potentially differ both systematically (bias) and in terms of imprecision (random error). The goal was to assess simultaneously the systematic differences among prototype OCT, OCT 1&2, Stratus, and Cirrus devices, and their corresponding imprecision standard deviations. Using the resulting calibration equations, consistent longitudinal thickness profiles were created for time periods as long as 17 years.

 

Methods

 

Global mean RNFL thickness measurements had been made so that some eyes were measured by both prototype OCT and OCT 1&2 (22 eyes), some were measured by both OCT 1&2 and Stratus (156 eyes), and others measured on both Stratus and Cirrus (92 eyes). No eye was measured simultaneously by more than 2 devices. A common factor structural equation model (SEM) was used to describe the measurement error. The unknown true thickness was represented as a latent variable. Two models were fit to the data: 1) simple model without image quality (IQ), and 2) enhanced model using IQ. Figure 1 shows the corresponding path diagrams. Model parameters were estimated using full information maximum likelihood (R language and environment for statistical computing with OpenMx package).

 

Results

 

Both simple and enhanced models showed similar scale parameter estimates with Cirrus and prototype showing smaller, and OCT 1&2 showing larger, scale parameters compared to Stratus. The larger the scale parameter, the smaller the measurement unit. Resulting calibration equations are shown in Listing 1. IQ had substantial and statistically significant effects when included in the enhanced model but with little impact on the device imprecision (except for OCT 1&2). As shown in Listing 1, device imprecision, after appropriate adjustment for scale differences was substantially higher for prototype and substantially lower for Cirrus.

 

Conclusions

 

Changes in device construction can impact the relative accuracy and precision of the resulting measurements, making it problematic for detecting overall trends in RNFL thickness over time. These problems can be alleviated to the extent that appropriate calibration techniques can be used to remove device artifacts due to systematic error. Calibration, however, has no effect on differences between devices in terms of imprecision.

 

 

 

Keywords: 552 imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)