Purchase this article with an account.
Li Tang, Xiaodong Wu, Kyungmoo Lee, Michael David Abramoff; Subvoxel Accurate Segmentation of 3-D Intraretinal Surfaces from SD-OCT Images by Non-Euclidean Deformation of the Graph. Invest. Ophthalmol. Vis. Sci. 2014;55(13):4794.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To automatically segment 3D intraretinal surfaces from spectral-domain optical coherence tomography (OCT) images with subvoxel accuracy by maximizing the utility of information extracted from intensity volume data using a non-Euclidean graph search (GS), which allows the use of lower resolution and cost scanners while still achieves comparable results as those from full scale segmentation on higher resolution data.
Five OCT volumes of 1536×61×496 voxels, covering a region of 9.37×7.93×1.92mm3 of the retina, were obtained from 5 normal subjects using a Spectralis (Heidelberg, Germany) OCT machine. Standard Iowa Reference Algorithm segmentation was performed on these volumes and were then down-sampled to 40×40×40 to assess the effect of standard and subvoxel GS. Two surfaces were segmented in the down-sampled volumes using conventional GS and the present method. By deforming the standard Euclidean graph using a displacement field obtained for each node from volume intensity data, we created a graph in non-Euclidean deformed graph space, in which the density of nodes increases at regions where certain transitions of image properties are more likely to occur. The upper envelope of the minimum closed set of displaced nodes corresponds to surface localization with subvoxel accuracy. Thickness of region bounded by the two coupled terrain-like surfaces was computed at 8000 (40×40×5) A-scans and compared quantitatively with results obtained at high resolution as ground truth.
The signed error of thickness was significantly decreased from 0.3411±0.5939 voxels using conventional GS to 0.2863±0.5099 voxels using subvoxel GS. The unsigned error was significantly decreased from 0.4842±0.4844 to 0.3480±0.4699. The percentage of A-scans with error larger than 0.5 voxel was reduced from 37.87% to 16.66%.
Our approach allows all standard GS techniques to work, thus retaining its advantages, such as global optimality and computational efficiency, with the same amount of nodes, memory and processing time. The identified surfaces provided a better representation of the smooth intraretinal tissue structure in presence of aliasing or partial volume effects introduced in imperfect imaging and digitalizing process. Thus more precise segmentation can be achieved on standard OCT, or lower cost imaging hardware can be used for the same results as current methods.
This PDF is available to Subscribers Only