Purpose:
To segment the retinal layers, especially the retinal nerve fiber layer (RNFL), on OCT-volume scans of normal subjects and glaucoma patients.
Methods:
Optic nerve head (ONH) centered volume scans were acquired using the Spectralis OCT (Heidelberg Engineering). The speckle noise of the volume data is reduced by weighted averaging in the 3D space. An edge detection along the A-scans yields an initial segmentation of five prominent layers excluding the outer nerve fiber layer boundary (ONFL). Model assumptions, that hold for glaucoma patients as well as for normal subjects, constrain the layer boundaries. An example for a model assumption is: The shape of a layer boundary should not differ from a RANSAC-fitted polynomial larger than a certain threshold. Finally, the ONFL is identified in between the inner plexiform layer boundary and the inner limiting membrane using an energy minimization approach (Mayer et al., Biomedical Optics Express, 2010). The algorithm was applied on 3 volume scans of normal subjects and 7 scans of glaucoma patients. An evaluation on the generated RNFL thickness maps was performed. The percentage of the scan area with an absolute thickness deviation of more than 0.01 mm from manually corrected segmentations is measured and defined as the segmentation error. An area around the ONH center with a diameter of 1 mm was excluded from the evaluation.
Results:
The average segmentation error on glaucoma patients is 11.0% compared to 4.4% on the normal subjects. The figure shows an automated segmentation result of a glaucoma patient and the corresponding color coded RNFL thickness map (red: thin, green: thick) overlaid on the SLO-image.
Conclusions:
The model assumptions made in our algorithm allow for an accurate segmentation of the retinal layers on OCT-volumes. Normal as well as pathologic data can be segmented with high accuracy.
Keywords: imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound) • image processing • retina