Detecting Glaucomatous Progression from Localized Rates of Retinal Height Changes in Parametric and Non-Parametric Frameworks

Abstract

Purpose: : To evaluate 3 new strategies that extend the statistic image mapping (SIM)¹ and detect glaucomatous progression from pixelwise rates of retinal height changes (PixR) with type I error controlled in parametric (PixR-P) and non-parametric frameworks (PixR-NP).

Methods: : In all techniques, the rate of retinal height change at each pixel was estimated using linear regression. Regression errors were assumed to be independent and identically distributed for non-parametric analysis (exchangeability criterion) and normally distributed with a constant variance over time for parametric analysis. 1) In PixR-NP cluster test (CT), familywise error rate (FWER) was controlled at both pixel-level (5%) and cluster-level (1%) using permutation tests. 2) In PixR-NP single-threshold test (STT), and 3) in PixR-P STT, k-FWER² was controlled at pixel-level (5%), respectively, using permutation tests and p-values (from t-tests) of the rate of change at each pixel. Permutation tests in PixR were conducted by permuting regression errors under the null hypothesis^3,4. For PixR-P and PixR-NP STTs, k was set to 5% of optic disk size to control k-FWER and an observed positive rate (OPR) was calculated as the ratio of number of pixels with significant rate of retinal height decrease (red-pixel) in disk to disk size. The criterion of progression was: for PixR-NP CT and SIM, p(cluster size) < 1% for at least one red-pixel cluster in disk; and for PixR-NP STT and PixR-P STT, OPR > 5%. Diagnostic accuracy was estimated using 267 eyes of 187 participants with at least 4 HRT-II exams from the UCSD Diagnostic Innovations in Glaucoma Study. 36 eyes progressed by stereophotos or visual field guided progression analysis; 210 patient eyes were non-progressing and 21 eyes were longitudinal normals.

Conclusions: : At 90% specificity, PixR strategies had better sensitivity than SIM. In contrast to SIM that controls false clusters only, PixR controlled both false locations and clusters of progression. Also, PixR strategies illustrated the use of k-FWER to control false positives in parametric and non-parametric analysis of rates of progression. The parametric PixR-P is computationally less intensive and thus, shows promise for wider adoptability in clinics.

Clinical Trial: : http://www.clinicaltrials.gov NCT00221897

Keywords: imaging/image analysis: clinical • optic disc • computational modeling