Fig 1.
Exemplary OCT B-scan from Spectralis SD-OCT showing eight intraretinal layer boundaries labeled as ,
…
.
Note that boundaries are delineated with red, yellow, magenta, white, cyan, green black and blue solid lines, respectively and the notations are summarized in Table 1. A parafoveal scan is chosen in order to show all the layers that are segmented by OCTRIMA 3D.
Table 1.
Notations for eight target boundaries, n denotes the frame number.
Fig 2.
OCTRIMA 3D framework to detect each intraretinal layer boundary using the shortest-path based graph search approach.
Fig 3.
The shortest-path based graph search methodology prefers a geometric straight line and may fail to delineate the ILM boundary in the central region of the fovea.
The ILM boundary can be detected correctly when the flattening operation uses the ILM border from a previous frame. Note that (a) and (c) are the raw OCT scans at the fovea and the resulting flattened image, respectively. (b) and (d) shows the results of the ILM boundary detection in (a) and (c).
Fig 4.
Illustration of the search region refinement using the inter-frame dependency approach.
Taking into account that the ILM boundary is delineated in the frame n − 1, the search region of could be limited to be
.
Fig 5.
Illustration of the biasing and masking operations for the boundary detection of the IPL-INL (red) and OPLo (blue).
The search region or ROI (red dotted rectangle) is the area between IS-OS and ILM (green solid lines). After element-by-element multiplication with lower bias map, the OPLo is more prominent in the gradient image and can be detected easily. A binary mask is generated to set all the pixels below OPLo to zeros and the second lowest boundary, IPL-INL, is detected using the shortest-path graph search.
Fig 6.
The overview of OCTRIMA 3D framework.
The boundaries labeled using blue and red fonts have the dark-to-bright and bright-to-dark transitions, respectively.
Fig 7.
Detection of A-scans affected by retinal blood vessels.
(a)The shadowing effect from the retinal blood vessels is more pronounced near the RPE (yellow rectangle, note the hypo reflective regions) and less pronounced near the ILM (green rectangle). (b) An example of the enface map of a macular volume OCT. (c) The result of RNFLo detection (blue solid line). The regions where the A-scans are affected by the blood vessel shadowing is highlighted with gray rectangles.
Fig 8.
Comparison of unsigned segmentation errors on six surfaces between Dufour’s algorithm (left column), the IOWA reference algorithm (middle column) and OCTRIMA 3D (right column) in the ETDRS regions.
The graph bar scale indicates the error magnitude in microns. The mean unsigned segmentation errors are reported in Table 2.
Table 2.
Comparison of average absolute detection error in unit of pixels and microns between Dufour’s algorithm, IOWA Reference Algorithm and OCTRIMA 3D in ETDRS region.
Fig 9.
The comparison between Dufour’s Software (magenta solid line), IOWA reference algorithm (blue solid line) and OCTRIMA 3D (red solid line) using manual labeling as the ground truth (green solid line).
Fig 10.
The comparison between OCTRIMA 3D (red solid line) and the algorithm by Chiu et al.
(blue solid line) using manual labeling as the ground truth (green solid line).
Table 3.
Comparison results between OCTRIMA 3D and the algorithm by Chiu et al. on Bioptigen OCT images. The error is quantified with (MSE± SSE, MUE, E95) in unit of pixels.
Table 4.
Comparison results between OCTRIMA 3D and manual labelings from two graders.
The manual labeling from Observer 1 is taken as the ground truth and the inter-observer difference is reported as a benchmark to evaluate the accuracy. The difference is evaluated using (MSE± SSE, MUE, and E95) in unit of pixels.
Fig 11.
Algorithms performance in the B-scan obtained from the patient with diabetic macular edema.
(a) The raw OCT B-scan. (b) The boundaries delineated by the built-in Spectralis SD-OCT software for the ILM and RPE-CH. The yellow arrows are indicating the boundary detection errors by the built-in software of the Spectralis device. (c) The boundaries delineated by OCTRIMA 3D for the ILM and the RPE-CH.
Fig 12.
The segmentation results obtained for the B-scan in the eye with dry age-related macular degeneration using Dufour’s software and the OCTRIMA 3D algorithm.
The legend of the boundaries is the same as Fig 1. (a) THe raw OCT B-scan. (b) The segmentation result of Dufour’s software. The IS-OS delineation failed at the left most and center area of the B-scan. (c) The initial segmentation results of OCTRIMA 3D detected retinal boundaries reliably except for the IS-OS in the drusen area (green doted line). By adjusting the flattening step, the IS-OS is delineated correctly (green solid line).