An Improved Method of Analyzing Polarimetery Measurements for Detecting Glaucoma: A Wavelet-Fourier Analysis Method Derived on GDx-VCC Polarimetry Data

Abstract

Abstract: : Purpose: Our prior efforts have sought to characterize the "double-hump" pattern of nerve fiber layer (NFL) thickness in terms of shape parameters derived from a Fourier analysis. Here we improve the mathematical model of NFL thickness using a new technique and test it on independent data sets. Methods: Glaucoma (defined by disc appearance) and normal eyes were scanned with a GDx-VCC (variable corneal compensator) polarimeter (Laser Diagnostic Technologies, Inc.) at several clinics. Thickness at 64 regions (disc diameter of 1.75) defined the double-hump curve. The spatial information of the curve was extracted by a Discrete Wavelet Transform (DWT) and then the concurrent detailed difference information was characterized by a Fourier transform (FFT). Dimensionality was then reduced by applying a Principal Component Analysis to the union of the DWT and FFT coefficients. A linear discriminant function of these blended FFT and DWT dimensions was used to classify eyes. The classification rule was derived and applied on separate samples (split half method applied 20 times and performance then averaged). Performance was determined in terms of sensitivity/specificity and ROC curve area and compared for this DWT/FFT/PCA method, the prior Fourier method (reported ARVO 2001), the manufacturer's GDx-VCC metric ("the nerve fiber indicator"), and other methods. Results: With the present sample (ca. 30 glaucoma patients and 50 normals), results were best with the new DWT/FFT/PCA method which yielded a Sensitivity/Specificity of .87/.98 and an ROC area of .97 on the training set and Sensitivity/Specificity of .84/.97, and ROC area of .96 on the independent testing sample. Conclusions: Excellent performance in discriminating between glaucoma patients and normals can be obtained with this mathematical technique of characterizing aspects of the shape of the NFL thickness curve.

Keywords: nerve fiber layer • imaging/image analysis: clinical