Abstract
Purpose: :
OCT as a non-invasive imaging technology enables the detection and follow-up examination of retinal pathologies such as retinitis pigmentosa. Today’s devices produce an enormous amount of data demanding an automatic assessment of relevant information such as layer boundaries, to do both supporting the ophthalmologist seeing degenerative changes but also to objectively quantify these degenerations. The aim of this study was the development of an automatic segmentation method to separate retinal layers and to quantify changes due to retinal degeneration in retinitis pigmentosa.
Methods: :
The retinal segmentation algorithm uses a graph theoretic approach to find optimal surfaces. The surface segmentation problem is transformed into computing a minimal s-t cut in a derived arc-weighted directed graph. For each layer individualized cost functions have been defined and optimized. We are using cost functions based on edge information and on shape information. The system is able to segment 5 individual layers (retinal nerve fiber layer; ganglion cell layer; inner plexiform/inner nuclear/outer plexiform layer; outer nuclear layer; interface of the inner and outer segments/retinal pigment epithelium/ bruch membrane) by identifying 6 boundaries. Each slice stack contains 49 slices with a resolution of 512x496 pixels for each slice. This results in rather complex graph setups. To handle this care has been taken to reduce the memory complexity and the computational complexity of the segmentation. The algorithm maintains individual sub regions of the slice stack with different spatial resolutions at the particular steps of the segmentation procedure.To evaluate the algorithm performance 14 OCT slice stacks have been used, seven healthy retinas and seven retinas showing symptoms of retinitis pigmentosa. Within this study we focused on the segmentation performance of the retinal layer between the interface of the inner and outer segments and Bruch membrane. The segmentation results have been compared with manual segmentation results performed by two senior Ophthalmologists.
Results: :
The comparison revealed ca. 96% conformity in the healthy group and ca. 89% conformity of the segmented layer in the group with pathologic slice stacks. An average segmentation took ca 2.39 min on a regular desktop PC.
Conclusions: :
The algorithm demonstrated both satisfying segmentation performance and execution time that would be suitable for a clinical setup.
Keywords: retina • image processing • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)