Evaluation of a Novel Proper Orthogonal Decomposition (POD) Framework for Detecting Glaucomatous Changes in Human Subjects

Abstract

 

Purpose:

 

To evaluate and compare a novel POD framework to Topographic Change Analysis (TCA) for detecting progression in HRT topographs.

 

Methods:

 

In POD framework, a set of basis vectors is constructed using an ensemble of optic nerve head (ONH) regions from topographs of an eye at baseline (BL). The bases capture any instrument inter/intra-session topograph variability and inherent variability in the ONH at BL. Changes in a follow-up topograph I_f are determined by reconstructing I_f using the BL bases. POD parameters a) Euclidean distance (L₂ norm), b) Image Euclidean distance and c) correlation between I_f and its reconstruction were computed and compared to TCA parameters a) total number of super-pixels with significant decrease in retinal height (red pixels), b) size of the largest red-pixel cluster (CSIZE) and c) CSIZE % of disc area, all within the disc margin. A total of 234 eyes with ≥ 4 HRT exams with a median follow-up of 5.5 years were selected from the UCSD Diagnostic Innovations in Glaucoma Study (DIGS) database. Standard Automated Perimetry Glaucoma Progression Analysis (GPA) was performed on all eyes using the 2 visual fields (VF) closest to the first HRT exam as baseline, and all reliable subsequent VFs as follow-up. GPA change was defined as ≥ 3 changed (p<5%) test points repeatable on ≥ 3 consecutive tests. For evaluation, HRT exams were categorized as changed and not-changed using the GPA VF change detection date.

 

Results:

 

Estimates of the Area Under ROC curve (AUC) for discriminating between changed and not-changed topographs and optimal parameter cutoffs are presented below. POD L₂ norm and TCA CSIZE resulted in the largest AUC in their respective frameworks and their difference (0.03) was statistically insignificant.  

 

Conclusions:

 

The new POD framework provides a similar performance as the TCA in detecting topographic changes. An advantage of the POD framework is its ability to detect and confirm changes using a single follow-up topograph by estimating instrument/structure variability from a set of BL topographs.

 

Keywords: imaging/image analysis: non-clinical • computational modeling