We have demonstrated the use of PS-OCT for an important application in quantitative retinal imaging: the segmentation of the RPE. Previous algorithms for segmenting retinal layers, including all algorithms that are used in commercial OCT instruments, are based on intensity images. While they can work well if retinal layers are just thickened or bended, they are based on certain assumptions like layer continuity. Therefore, these algorithms frequently have problems if layers are interrupted, discontinuous, or partly missing. E.g., they frequently fail in identifying RPE atrophies, taking Bruch’s membrane for the RPE in atrophic areas. The two segmentation algorithms presented here are the first that are not based on backscattered intensity but on polarization properties of backscattered light, i.e. they exploit an intrinsic tissue property. A main advantage is therefore that these algorithms do not rely on any assumptions about the order of layers and do not require that a layer is contiguous. Layer interruptions, e.g. by atrophies or shadowing by blood vessels, do not disturb the algorithm, the RPE is found on both sides of such interruptions.
We demonstrated two different algorithms: a simpler algorithm based on the retardation data, and a more sophisticated algorithm based on averaged Stokes vector elements. While the first algorithm uses only the amplitudes measured in the two polarization channels, the latter also takes the phase difference of the signals into account. Since there is more information available in the latter case, it is worthwhile to consider if this algorithm can provide superior results. Although both algorithms perform very similar in simple cases (healthy eye, ) there are cases where the Stokes vector based algorithm is superior.
shows a B-scan through the red patch within the RPE atrophy visible on the left side of . Both segmentation algorithms are compared. shows the results obtained from algorithm 1 (retardation SD), while in algorithm 2 (Stokes vectors) was used. shows the corresponding retardation image, a direct comparison of the segmentation results provided by the two algorithms. Although the two algorithms yield rather similar results in most areas, there are also differences. In the region below the RPE atrophy, where light, due to the absence of the RPE, penetrates down to the strongly birefringent sclera, the algorithm based on retardation SD falsely segments some regions in the sclera (indicated with a blue circle in ). The strong retardation in this area obviously fools the algorithm by increasing the width of the retardation SD (which causes the red patch mimicking a thickened RPE in ). The algorithm based on Stokes vectors is more reliable in this case. Therefore, there is a trade-off between algorithm speed and reliability. Optimized routines written in a more efficient programming language and parallel processing will overcome this problem and the Stokes vector algorithm might be favored for further work.
Fig. 10 B-scan of RPE atrophy: comparison of segmentation algorithms. (a) retardation SD based algorithm; (b) Stokes vector based algorithm; (c) retardation image; (d) Comparison of segmentation algorithms. Blue: data points segmented by both algorithms; red: (more ...)
An interesting feature of our method is the possibility to quantify the thickness of the pigmented tissue. However, the interpretation of the thickness maps (cf. Figs. and ) deserves some caution. The RPE consists of a cell monolayer of only a few μm thickness. The layer thickness obtained by the segmentation process is larger than the value expected for the thickness of the healthy RPE. The reason is that the segmented structure is the result of a convolution of the real tissue extension with the evaluation window function. Deconvolution might improve thickness measurements, however, would require a deeper knowledge of the statistics of the polarization state distribution caused by the RPE. There is a trade-off between the number of data points available for retardation and Stokes vector element statistics (and thus the reliability of the algorithms) on the one hand, and achievable resolution on the other hand. The window size presently used (~70 μm (x) × 18 μm (z)) is a compromise. Improvements might be achieved by improving the depth resolution (larger source bandwidth) and/or by 3D windowing (which would require a larger sampling density in y-direction), providing more data points for histogram calculations and/or smaller in-plane window sizes and thus better resolution of segmented RPE.
Although the thickness of the normal, healthy RPE cannot be accurately measured, the thickness maps are valuable tools for detection of RPE loss or of accumulations of depolarizing tissue to a thickness well beyond the evaluation window extension (cf. ). None of the commercial OCT instruments presently available can detect such accumulations, and no gold standard for any measurements of pigmented tissue thickness presently exists.
The algorithms presented here might be useful for different tasks of quantitative evaluation of the human retina. On the one hand, mapping of total retinal thickness might be improved. Recent studies with intensity based OCT have shown that errors of retinal thickness measurement were common, with at least some error present in 92% of scans [34
]. The inability to clearly identify the RPE in intensity based OCT images makes it rather difficult to create a reliable algorithm for automated retinal thickness measurements. The segmentation of the RPE via its polarization properties might therefore improve retinal thickness measurements. On the other hand, thickening of the RPE can be visualized and the presence and size of atrophies can be quantified with our method and displayed in en face maps. This en face visualization of the RPE can then be compared with other en face imaging methods, like fluorescein angiography or auto fluorescence imaging, and perhaps help to avoid invasive imaging methods at least in some cases. Another important application of RPE segmentation could be automated screening of the ocular fundus for small RPE atrophies that are presently very difficult to detect. These lesions could be important for the treatment decision and outcome prediction of modern antiangiogenetic therapy in neovascular AMD, where the presently evaluated morphometric parameters like retinal thickness are known to be unreliable predictors [35
]. A reliable quantification of these RPE lesions would be especially important for follow up studies and treatment control.