Automatic count of outer retinal hyper-reflective bands in normal and retinitis pigmentosa subjects

Abstract

 

Purpose

 

Investigation of structural changes in the outer retina in normal and retinitis pigmentosa (RP) subjects requires segmentation of the individual retinal layers. For segmentation, the number of discernible hyper-reflective bands needs to be determined. Here, we present and evaluate a method to count the number of hyper-reflective bands in the outer retina in normal and RP macular OCT volumes.

 

Methods

 

The number of discernible hyper-reflective bands in the outer retina may vary due to disease or overlapping bands. The bands were modelled by a Gaussian mixture model (GMM) with a varying number of components. Model parameters were optimized by maximum-likelihood estimation. The information complexity (ICOMP) criterion was then used to select the best model and consequently the number of bands was set to the selected model's number of components. ICOMP balances the quality-of-fit with the interdependence of the parameters. A complex model is only preference if it was needed to model the data, i.e. if the better fit is not penalized due to model degeneracy.<br /> <br /> To validate the detection, 2 normal and 2 RP patients were imaged with a Spectralis OCT system (Heidelberg Engineering, Germany). The macular scan protocol combined 37 B-scans composed of 512 A-lines of 496 pixels, with 35 times averaging. A medical doctor manually annotated the number of bands visible in the outer retina on every second B-scan.<br /> <br /> Manual annotations were compared to automatic counts and the precision and sensitivity were calculated per A-line.

 

Results

 

Examples of manual and automatic band counts are shown in Figure 1. The results of a quantitative assessment are listed in Table 1. The precision for one, two and three bands was 88, 81 and 88%, respectively. Further, the corresponding sensitivities were 95, 84 and 77%, respectively.

 

Conclusions

 

A method for counting the discernible outer retinal bands was presented. The preliminary results indicate that the method was able to correctly count the visible bands in over 80% of A-lines. Further improvements of the results may be obtained by post-processing.  

 

Figure 1: Examples of manual and automatic counts of the visible hyper-reflective bands in a macula-centered en-face image for a normal and an RP subject.

 

 

Table 1: Confusion matrix of manual and automatic counts of visible bands (one, two and three). The numbers in the table correspond to the number of A-lines in all eyes.