Comparative Assessment of OCT Segmentation Metrics between Manual Corrected Image Processing and Machine Learning in Severe Inflammatory Disease and Retinal Dystrophy

Abstract

Purpose : To determine the differences in optical coherence tomography (OCT) segmentation metrics between machine learning based assessment and an expert-reader corrected logic based algorithm in eyes with severe outer retinal diseases. A machine learning based segmentation model created from normal, diabetic retinopathy and age related macular degeneration eyes was assessed in this experiment. OCT images from severe outer retinal diseased eyes (uveitis patients with syphilitic retinitis and retinitis pigmentosa) were analyzed and compared to a manually corrected expert readers.

Methods : 21 eyes were included in this experiment (9 RP and 12 syphilitic). For this analysis, the inner limiting membrane (ILM), outer nuclear layer (ONL), ellipsoid zone (EZ), and retinal pigment epithelium (RPE) were segmented using both deep learning and a logic based algorithm (LBA) with expert reader (ER)-corrected segmentation. The intraclass correlation coefficients (ICC) for ER-corrected LBA and machine learning were calculated and compared using segmentation metrics for central subfield thickness and central subfield volume. Intraclass correlation coefficients were created (MS-Excel, RealStatistics 2020) with 95% confidence intervals.

Results : ICC scores were highest for total retinal thickness measurements including central subfield thickness (.619) and central foveal mean thickness (.826). Machine learning displayed poor correlation when compared to ER-corrected LBA for all EZ based measures including EZ/RPE central subfield thickness (ICC 0.062) and central subfield volume (ICC 0.059). Enface based measures including maps detailing zero micron thickness and less than 20 micron thickness also had poor correlations. When separating out by disease subtype, there were no differences between correlations in syphilis retinopathy nor in RP.

Conclusions : In this machine learning model created from normal, AMD and DR pathologies, there was poor correlation to expert readers when assessing EZ based metrics in eyes with severe outer retinal pathology. Though our numbers are small, specific models created from manually graded severe retinal pathology may be needed to improve performance when assessing these complicated eyes. When assessing severe EZ pathology, our results suggest that manual grading is still needed.

This is a 2021 ARVO Annual Meeting abstract.