Methods for manual and automated detection of the four outer retinal hyper-reflective bands in optical coherence tomography scans

Abstract

Purpose: To develop and validate methods to extract position and thickness for the four outer retinal hyper-reflective bands in spectral domain optical coherence tomography (SD-OCT) scans of normal and diseased eyes for comparison to structures visible in histological studies. An automatic method of evaluation is desirable to quickly and consistently generate measurements. To provide a high level of confidence in the results the automated method should be validated using multiple manual methods and multiple observers and patients.

Methods: Our previously reported method of measuring band position and thickness in SD-OCT image data models each band as a Gaussian. The bands overlap. An optimization method extracts from the resulting sum of Gaussians the center position, width, and amplitude for each band. We have implemented two different manual segmentation procedures. Comparisons among these methods are being used to validate the automatic method. The validation study has access to over 500 Spectralis SD-OCT volumes from the UAB ALSTAR study. The first manual method requires the observer to sketch smooth curves. It allows for global evaluation of the image. The second manual method focuses on discrete measurement locations. It allows the observer to closely examine local regions in the image. For both manual methods the user works with contrast enhanced images designed to mimic the display produced by the Spectralis. The automatic method uses raw image data. Three observers have been trained. Two have prior professional experience segmenting SD-OCT scans. Intra-class correlation coefficients and Bland-Altman plots will be used for the statistical analysis.

Results: Three different segmentation methods were created, all suitable for characterizing the four outer retinal hyper-reflective bands in SD-OCT images. Preliminary comparisons of the manual and automatic methods show good agreement between band center locations and some (systematic) differences between thickness measurements that may be attributed to difficulties in judging gradients on displayed images.

Conclusions: The use of multiple methods provides an opportunity for thorough evaluations of both automatic and manual methods. Feedback on user interface is being incorporated into new versions.

Keywords: 551 imaging/image analysis: non-clinical • 688 retina