Accuracy of algorithmic segmentations of ILM and RPE against manual grading

Abstract

Purpose : Optical coherence technology (OCT) provides the ability to measure retinal thickness and visualize the various tissues in the retina. Retinal thickness is measured as the distance between the internal limiting membrane (ILM) and the retinal pigment epithelium (RPE). We have developed a new prototype multilayer segmentation algorithm (MLS) that segments the ILM and RPE as well as other layers. The purpose of this study is to compare the performance of the prototype and commercial algorithms against manual segmentation.

Methods : Accuracy of the commercial and MLS algorithms was compared against ILM and RPE segmentations hand-drawn by graders trained in segmenting OCT volumetric images (gold standard). Eyes were scanned with the 200x200 and 512x218 macular cube scans of the Cirrus^TM HD-OCT 4000 (ZEISS, Dublin, CA). Subjects with and without retinal diseases were recruited from 4 study sites. ILM and RPE were segmented on central B-scan. A total of 284 200x200 scans and 303 512x128 scans were analyzed.

Results : Fig 1 shows a central B-scan with commercial, multi-layer, and gold standard segmentations overlaid in different colors. For the 200x200 macular scan (Table 1), the mean difference of segmentation between the gold standard and the commercial segmentation was 7.35µm for ILM and 7.17µm for RPE; and between gold standard and MLS was 3.47µm for ILM and 9.32µm for RPE. For the 512x128 macular scan, the mean difference between the gold standard and the commercial segmentation was 7.25µm for ILM and 7.12µm for RPE; and between gold standard and MLS was 3.02µm for ILM and 9.46µm for RPE. The mean thickness difference for commercial segmentation was 13.25µm for 200x200 scan and 15.02µm for 512x128 scan. The mean thickness difference for MLS was 14.19µm for 200x200 scan and 16.33µm for 512x128 scan.

Conclusions : The prototype MLS algorithm performs similarly to the commercial segmentation, while providing the possibility of segmenting additional layers. Differences in both algorithms with respect to gold standard are influenced by anatomic variations due to retinal diseases, scan type and quality, and individual preferences of trained experts.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.

 

Fig 1: B-scan of a retina with macular degeneration and Drusen, showing ILM and RPE segments for gold, commercial algorithm and MLS

 

Table 1: ILM and RPE segments for commercial and MLS algorithms against gold standard