Purchase this article with an account.
M. Niemeijer, M. Sonka, M. K. Garvin, B. van Ginneken, M. D. Abramoff; Automated Segmentation of the Retinal Vasculature in 3D Optical Coherence Tomography Images. Invest. Ophthalmol. Vis. Sci. 2008;49(13):1832.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evaluate an automated computer algorithm for segmentation of the retinal vasculature in 3D spectral OCT. Segmentation of vessels will allow inter-subject registration of multiple 3D-OCT volumes over time, registration of OCT with fundus photographs and angiograms.
28 3D-OCT volumes were obtained (Carl Zeiss Meditec Cirrus 3D-OCT, 200x200x1024 voxels) from 28 normal subjects (28 eyes), 14 were centered on the optic nerve head (ON) and 14 were centered on the fovea. An automated layer segmentation algorithm  was applied to each volume to identify individual retinal layers. The vessel silhouettes were used to identify the vessel positions. Using the layer segmentation, the outer retinal layer which showed the highest vessel silhouette contrast was averaged in the z direction. A 2D projection image of the vessel silhouettes was obtained for all 3D OCT volumes. To develop a method for vessel segmentation based on the projection images, the set of 28 images was split into a training set of 12 images (6 ON, 6 macula-centered) and a test set of 16 images (8+8). For each of the training images, all vessel pixels were indicated by an expert. Next, two separate kNN classifiers were trained for the ON-centered and the macula-centered projections. A Gaussian derivatives filterbank (2nd order, σ=1,2,4,8) was applied to each of the 2D projection images to extract pixel features. All pixels from the training images were used in the training procedure. The two trained classifiers were applied to their respective test sets. A human observer indicated for 100 randomly selected pixels in each of the test images (1600 pixels in total) whether the pixel belonged to a vessel or not.
Comparison of the automated segmentation with the observer-generated independent standard indicated an area under the ROC curve (Az) of 0.96 for both the ON and macula-centered volumes combined. For the ON-centered test images, the Az was 0.97 and for the macula-centered images the Az was 0.94. As the results indicate, smaller vessels in the macula-centered images were more difficult to segment than the larger vessels near the ON.
A novel automated algorithm can segment the retinal vasculature in 3D-OCT images with a performance close to that of a human expert. M. Haeker, M. Sonka, R. Kardon, V.A. Shah, X. Wu, and M.D. Abramoff, "Automated segmentation of intraretinal layers from macular optical coherence tomography images", Medical Imaging, 6512, SPIE, 2007
This PDF is available to Subscribers Only