Progression Detection in Simulated Scanning Laser Polarimetry Images

Abstract

 

Abstract:

 

To determine the sensitivity and specificity of several related progression detection methods for glaucoma based on real and simulated Scanning Laser Polarimetry images.

 

One eye of 24 normal subjects was imaged three times on four days with a GDx VCC (with ECC). Seven event–based progression detection methods were tested. These methods, based on one or two baseline visits and one to three follow–up visits, related the observed change to the measurement repeatability. They were trained on the real images for a fixed specificity of 97.5%. The sensitivity of these methods for various reproducibilities was determined by a simulation study for either localized (in a 20 degrees sector) or diffuse loss. The reproducibility of measurements of the eyes was modeled as the mean standard deviation across the image.

 

The minimally detectable loss, defined as the minimum loss that resulted in a sensitivity of 90%, for each method is shown in the figure. The first number in the legend shows the number of baseline visits. The second number is the number of follow–up visits and the last two numbers denote the number of passed/total tests to indicate change (e.g., 2/3 means that 2 out of 3 tests have to indicate loss for the eye to be classified as showing significant change).

 

The minimally detectable loss decreases if more visits are available. Likewise, a larger number of required successful tests results in a lower minimally detectable loss, with the exception of tests with three follow–up visits.

 

Progression detection seems feasible when multiple images are acquired per visit, thereby allowing the estimation of the reproducibility of the eye. Depending on the area showing loss, detection of a loss of typically 3 to 15 µm is possible with a specificity of 97.5% and a sensitivity of 90%.

 

 

 

Keywords: computational modeling • image processing • nerve fiber layer