Purchase this article with an account.
David Stein, Nadia Waheed, Mehreen Adhi, Salim Semy, Jay Duker, Walter Kuklinski; Automated Detection of Hyper-reflective Structures Associated with Diabetic Retinopathy using Spectral Domain Optical Coherence Tomography (SD-OCT). Invest. Ophthalmol. Vis. Sci. 2013;54(15):5525.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The objective of this study was to determine if automated algorithms applied to SD-OCT macular images could identify hard exudates (HE) and microaneurysms (MA).
SD-OCT macular image studies were performed at the New England Eye Center using a CirrusTM 4000. Clinicians at the New England Eye Center classified each of the training and test images. The training images consisted of 64 scans from normal eyes. The test images consisted of 26 scans of eyes with diabetic retinopathy (DR) but no HE or MA and 40 scans from eyes with DR exhibiting either HE or MA or both. A suite of layer segmentation algorithms were applied. A local image normalization algorithm that enhances locally bright pixels was developed and applied to the region of the retina from the inner to the outer limiting membranes. Three-dimensional clusters consisting of contiguous locally bright pixels were subsequently identified using a pattern classification method based on features that included size, maximal normalized brightness, shape, and location. Probability distribution functions (PDFs) for the features were estimated from the healthy training data. An anomaly detection statistic was calculated for each cluster. Detection statistics based on the number of anomalous clusters in the ganglion cell, the inner nuclear, and the outer plexiform layers were used to identify DR. A sum detector aggregated the number of clusters in these layers, whereas the outer plexiform (OP) detector was based on the number of clusters in the OP layer.
Figure 1 shows a slice of a locally normalized macular OCT image on an eye that clinicians confirmed contained hard exudates. The red boxes in Figure 1 surround anomalous hyper-reflective clusters identified by the automated detection algorithm. Figure 2 shows the performance curves for the classifier algorithm. At a PFA of 12% the sum detector achieves a PD of 90% while the OP detector achieves a PD of 84%. At a PD of 100%, the OP detector has a PFA of 35%, while the sum-detector has a PFA of 58%.
Automated analysis of macular OCT images can be used to distinguish eyes containing HE or MA from those that do not. This approach can be used with our previously developed automated methods to identify macular edema in a comprehensive automated screening tool for all eyes with non-proliferative diabetic retinopathy.
This PDF is available to Subscribers Only