Specificity of Trend-based Progression Analysis (TPA) to Detect Progression in the Ganglion Cell Inner Plexiform Layer (GCIPL) Thickness Map

Abstract

 

Purpose

 

To evaluate the specificity of a false discovery rate (FDR)-controlled, TPA to analyze progressive GCIPL thinning in individual superpixels (SPs) of serial GCIPL thickness maps in normal eyes.

 

Methods

 

Cirrus HD-OCT Macular Cube data of 25 eyes from 25 normal subjects followed weekly for 8 consecutive weeks were analyzed. For each eye, GCIPL thickness maps were extracted and aligned to the first visit using the instrument’s segmentation and blood vessel-based registration algorithms. The analyses were restricted to the SPs (created by averaging 4x4 pixels) located in the annular ellipse centered on the fovea used in the instrument’s Ganglion Cell Analysis module. Individual SP data were analyzed using linear regression (LR). To reduce the probability of Type I error due to multiple testing, the significance level of testing in each SP was determined after controlling the FDR at ≤5% by a two stage method [1, 2]. Progressive thinning in a SP was encoded in yellow in the change map if a significant negative trend were found with P≤5% in an individual LR analysis, and encoded in red if a significant negative slope were detected after controlling the FDR at ≤5%. Progression in each method was defined where ≥20 adjacent SPs were encoded. Any progressions were assumed to be false positives for these normal eyes. Visits ≥3 were considered follow up visits.

 

Results

 

6 (24%) and 0 (0%) eyes showed progression detected by LR and TPA at the final visit, respectively, with almost no SPs flagged for change by TPA. The mean total of SPs flagged and the specificity of each method at each of the 6 visits are shown in Table 1. Figure 1 shows an example of TPA with 1 progressing SP versus 59 progressing SPs for LR.

 

Conclusions

 

TPA with FDR control achieved higher specificity than standard LR for detecting GCIPL thinning in the SP change maps for the normal eyes. By reducing Type I error, localized rate of change maps for individually progressing SPs may be enabled by TPA and provide an adjunct approach to analyze progressive GCIPL thinning in conjunction with event-based methods.<br /> <br /> [1] Benjamini et al., 2001. Annals of Stats, 29:1165-1188. [2] US Pub No. 2013/0308824  

 

Table 1. Mean superpixel progression and overall specificity at each follow up visit for LR and TPA.

 

 

Figure 1. TPA with FDR control (red) improves specificity versus standard LR (yellow).