Comparison of multi-retinal layer segmentation vs. existing segmentation in CIRRUS OCT images

Abstract

Purpose : Segmentation of the inner retinal layers (IRL) in optical coherence tomography (OCT) images is an important tool for accurate diagnosis of eye diseases. The existing segmentation algorithm (ESA) in CIRRUS estimates the positions of the inner plexiform layer (IPL) and outer plexiform layer (OPL) based on the internal limiting membrane (ILM) and retinal pigment epithelium (RPE). To improve the accuracy of the segmentation of these layers, we employ a multi-layer segmentation algorithm (MLS) which truly segments layers instead of estimating their position. Here we compare the accuracy of MLS and ESA to manual segmentation.

Methods : Subjects with normal and diseased eyes were imaged with CIRRUS™ 6000 AngioPlex (ZEISS, Dublin, CA). Diseases included age-related macular degeneration, diabetic retinopathy, vitreoretinal interface abnormalities, and other diseases. OCT volumes were acquired using the 3x3 mm, 6x6 mm and HD 6x6 mm, 8x8 mm, HD 8x8 mm and 12x12 mm angiography scan types. In five B-Scans from each volume (at 10%, 30%, 50%, 70% and 90% of the cube width), ILM, IPL, OPL and RPE were segmented by MLS, ESA and two human graders using a semi-automatic OCT segmentation tool (EdgeSelect). All segmentation lines were manually checked for accuracy by the graders. The ESA used the following estimations for the depths of the IPL & OPL: Z_OPL = Z_RPE - 110μm, and Z_IPL = Z_ILM + 0.7 (Z_ILM - Z_OPL).
Bland-Altman (BA) plots were generated comparing the mean of the two manual segmentations with the results by ESA and MLS over all five groups (normal and diseased) and for the following layers: ILM, IPL, OPL and RPE.

Results : Table 1 contains the 95% agreement interval metrics for 290 B-Scans divided into normal: 75, DR: 45, AMD: 55, VRI: 55 and other: 60. MLS shows smaller mean differences to manual than ESA for all subject groups and layers. The MLS-to-manual agreement interval in the BA plots is also less than half as wide as the ESA-to-manual agreement interval over all groups for the IPL and OPL. This indicates that the MLS’ segmentations follow the manual grading, i.e. the gold standard, more closely than the segmentations generated by the ESA algorithm, especially for these IRL (see Figure 1).

Conclusions : MLS segments the IRL more accurately than the ESA and performs similarly for the ILM and RPE, albeit with smaller bias, showing that MLS is an improvement over ESA.

This is a 2020 ARVO Annual Meeting abstract.