Automated Intra-Retinal Layer Thickness Profile Analysis in Optical Coherence Tomography Images by Iterative High-Pass Filtering

Abstract

 

Purpose

 

To develop an automated system that segments seven retinal sub-layers in OCT images of both normal eyes and eyes with diabetic macular edema (DME).

 

Methods

 

Heidelberg Spectralis OCT images were used. A total of 87 SD-OCT macular line scans were used with 22 of the scans containing significant disruptions of the retinal micro-structure due to cysts. For the algorithm each OCT image was subjected to automated speckle noise removal using both a Fourier-domain based error metric and a Wiener deconvolution algorithm. Next, an iterative multi-resolution filtering algorithm was used to detect the next most significant layer using high-pass filtering. Finally, the automated thickness profiles for each intraretinal layer were compared with manually marked ground-truth.

 

Results

 

We observe that for the images with no retinal disorganization, the mean and standard deviation of error in μm for segmenting each retinal layer is as follows: NFL (10.23±2.27), IPL (22.68±1.74), INL (16.67±4.3), ONL (6.61±4.03), PIS (2.29±0.61), PE (7.29±3.17). For images with significant retinal disorganization, these errors are as follows: NFL (6.07±2.53), IPL (32.68±4.23), INL (20.59±3.17), ONL (13.68±4.13), PIS (3.729±1.91), PE (11.07±2.8). The mean measured intra-retinal thickness profiles in μm are as follows: Outer layers (106.09± 8.25), Inner layers (86.50 ±4.89), INL (38.11±5.15), ONL (65.42±4.94), Inner/Outer segments (38.95±2.92). Finally, the mean retinal layer thickness using the automated algorithm is compared to the manual segmentation results using the correlation coefficient (r) and the R² metric as follows: Outer layers (r=0.86, R² =0.74), Inner layers (r=0.74, R² =0.54), INL (r=0.90, R² =0.81), ONL (r=0.97, R² =0.94), Inner/Outer segments (r=0.91, R² =0.82). The complete segmentation algorithm requires less than 35 seconds per image on a 2.53GHz Intel Core i3, 2GB RAM system.

 

Conclusions

 

This novel system can reliably measure intraretinal sub-layer thickness in both normal eyes and in eyes with DME. These measurements may help characterize patient status and predict prognosis.

 

 

Automated segmentation of retinal layers. The retinal layers are represented as: RFNL (Blue), IPL/GL (Cyan), INL (Red), ONL (Yellow), PIS (Magenta), POS (Black), PE (Green).

 

Keywords: 551 imaging/image analysis: non-clinical • 496 detection • 552 imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)