Abstract
Purpose: :
The retinal surfaces obtained from the OCT scanners are far from flat, and show different forms of distortions in the B-scan and C-scan slices. These artifacts are thought to be the result of a number of factors such as the corneal curvature, motion of the eye and the positioning of the camera. Flattening the dataset makes visualization easier by bringing the dataset into a more consistent shape, which also allows for the efficient truncation of the dataset. Here, we present a quantitative comparison of two automated methods that eliminate distortions characteristic to optical coherence tomography images.
Methods: :
First, the retinal surfaces are detected through an automated 3-D graph-theoretic approach. A surface is then selected and used to determine the flattening plane through two consecutive thin-plate spline fits. The first spline-fit uses a smoothing regularization term and an equal number of control points in both axial directions to approximate the distortion seen in B-scans. For the second spline-fit, a smaller regularization term and a larger number of points in the direction of the slice acquisition is used to approximate the rippling seen in C-scans. In both stages, points within the neural canal are avoided using a circular mask to approximate the region. This two-stage approach is compared with the method that uses a single spline-fit.The method is quantitatively validated using depth maps of the optic nerve head constructed from fundus photographs (Tang ’10, SPIE) taken at two slightly different angles. Since the depth images approximate the shape of the optic nerve head, they can be compared with the top surface of the flattened dataset. The normalized depth map is registered to the OCT dataset and the mean unsigned difference is computed within valid areas of the depth map (excluding the neural canal).
Results: :
Over 30 glaucomatous datasets, the mean unsigned difference between the depth maps and the top surface from the datasets flattened (which are also normalized consistently) by the two-spline flattening approach was significantly smaller than the single spline-fit (0.215 ± 0.056 and 0.127 ± 0.039, respectively; p < 0.001).
Conclusions: :
Although various methods exist for flattening OCT datasets, thus far none have been quantitatively validated. The presented two-spline flattening method was compared quantitatively with a single-spline approach and found to be more robust.
Keywords: image processing • optic disc