Current retinal segmentation algorithms are based on layer or layer-boundary delineation. Although it works well for cases without sever structural abnormality, its applicability to heavily damaged tissue is low.
In this study, we demonstrate pixel-wise segmentation of retinal pigment epithelium (RPE) and choroidal stroma by using multiple contrasts obtained by Jones-matrix OCT (JM-OCT). The method can be applicable equally to normal and pathologic cases.
A custom made posterior JM-OCT is used to obtain multi-contrast retinal images. A single scan of JM-OCT provides multiple images including scattering OCT, attenuation coefficient, birefringence, degree-of-polarization uniformity, and OCT angiography.
The tissue segmentation is done by applying a threshold to a “feature” which is a synthesized from the multi-contrast images. For RPE segmentation, the feature is defined as the third order combination of the optical contrasts. Then, the pixel is classified as RPE if the feature is larger than a particular threshold value. For choroidal stromal segmentation, another feature is defined as a second order combination of the optical features.
After segmenting the RPE and choroidal stroma, RPE thickness map, RPE elevation map, whole choroidal thickness map, choroidal stromal thickness map, and choroidal vessel/stromal ratio map are generated in a normal and a myopic CNV complicated with non-sunset Vogt-Koyanagi-Harada disease (VKH-mCNV) cases. The generated maps of VKH-mCNV case visualize the absence of RPE, thinner choroidal thickness than normal case, and CNVs.