Abstract
Purpose :
The Notal OCT Analyzer (NOA) is a Deep leaning-based algorithm developed for Notal’s Home OCT System. This study evaluated the performance of the NOA in segmentation of intraretinal fluid (IRF) and subretinal fluid (SRF) regions, and quantification of fluid volumes, in cube scans from self-imaging of subjects with neovascular age-related macular degeneration at home with the Home OCT System
Methods :
The data comprised a total of 2024 B-scans, with 8 eyes contributing 3 volume scans each. Each B-scan was segmented independently by 3 human expert graders to identify and quantify IRF and SRF. Intra-group and inter-group agreement were compared (i.e., agreement between the NOA and human graders, versus within the human graders). The outcome was the Dice coefficient for segmentation comparisons (pixel-wise and compared by Bonferroni-adjusted Kruskal-Wallis test), and the Pearson correlation coefficient and absolute mean difference for volume comparisons.
Results :
Of the 2024 B-scans, the number identified with fluid by Grader 1 was 449 (356 with SRF and 96 with IRF). For the segmentation, the Dice coefficient of the NOA was not statistically different from that of the human graders. For the fluid quantification, the correlation coefficient between the NOA and the human graders (mean of three) was . The absolute mean difference between the NOA and the human graders (mean of three) was 1.2 nL. This was lower than the difference between Graders 1 and 2 (1.5 nL) and Graders 1 and 3 (1.8 nL), and similar to the difference between Graders 2 vs 3 (1.2 nL).
Conclusions :
The NOA performs robustly in automatically segmenting and quantifying retinal fluid from cube scans acquired by patient self-imaging using the Home OCT System. In sample of eyes, the agreement with human expert grading was very high, and was similar or superior to the level of agreement between different human experts. Automated analysis of OCT images is a prerequisite for efficient remote patient monitoring with a Home OCT System that generates data at scale over time and may be useful for patient care.
This is a 2021 ARVO Annual Meeting abstract.