Automated 3-D Retinal Layer Segmentation of Macular OCT Images with Retinal Pigment Epithelial Detachments

Abstract

 

Purpose

 

Automated retinal layer segmentation of OCT images has been successful for normal eyes but remains an open problem for eyes with retinal diseases. We present a method to automatically segment the retinal layers in 3-D OCT data with retinal pigment epithelial detachments (PEDs), which is a prominent feature of many chorioretinal disease processes.

 

Methods

 

Macular-centered SD-OCT (Topcon, 512×64×480 voxels, 11.72×93.75×3.50µm3) of 9 eyes diagnosed with PEDs were acquired. Automated 3-D retinal layer segmentation based on graph search was achieved as follows: First, surfaces 1-6, which were not dramatically affected by PED, and a surface combined by surface 7 and 10 (surface 7’) were detected by a multi-resolution approach[1]. Then the floor of RPE layer (surface 11) was detected below surface 7’ with large smoothness constraints to allow the abrupt change caused by PED. A surface (surface 12) indicating the normal RPE floor was detected using the same cost function but small smoothness constraints. Subsequently, the PED footprints indicating which A-scans were associated with PED were obtained by calculating the distance between surface 11 and 12. Finally, the OCT volume was flattened using a reference surface combined by surface 11 and 12, which removed the PED area and regained the normal appearance of surfaces 7 and below. In this flattened image, surface 7-10 were detected with corrections around the PED area.

 

Results

 

Manual segmentation was performed by two experts for surfaces 1,2,4-7,10 and 11, which were discernible to human eyes. For each OCT volume 10 B-scans (every sixth among the total 64) were selected for manual tracing, and the average z-position of the two results were used as the reference standards. The overall mean unsigned border error of the automatic segmentation was 2.25±0.69 voxels (7.89±2.41µm) and was comparable to the inter-observer variability 2.15±0.53 voxels (7.54±1.87µm).

 

Conclusions

 

Automated 3-D retinal layer segmentation of macular OCT images with PED has been achieved. Further analysis on retinal layer thickness, intensity and PED area/volume can be applied based on the segmentation result, which may help the diagnosis of various diseases related to PED.

 

 

Figure 1: Segmentation result. (a) A (truncated) B-scan of macular OCT volume with PED (b) Image (a) overlaid with twelve retinal surfaces obtained by automated segmentation.

 

Keywords: 550 imaging/image analysis: clinical • 701 retinal pigment epithelium