Premium
Reproducibility of pixel‐by‐pixel analysis of Heidelberg retinal flowmetry images: the Thessaloniki Eye Study
Author(s) -
Mavroudis Leonidas,
Harris Alon,
Topouzis Fotis,
Wilson M. Roy,
Yu Fei,
Anastasopoulos Eleftherios,
Koskosas Archimidis,
Siesky Brent,
Pappas Theofanis,
Founti Panayiota,
Coleman Anne L.
Publication year - 2008
Publication title -
acta ophthalmologica
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.534
H-Index - 87
eISSN - 1755-3768
pISSN - 1755-375X
DOI - 10.1111/j.1600-0420.2007.01003.x
Subject(s) - pixel , wilcoxon signed rank test , mathematics , reproducibility , intraclass correlation , standard deviation , medicine , statistics , ophthalmology , mann–whitney u test , artificial intelligence , computer science
. Purpose: To identify the investigator effect in the analysis results of Heidelberg retinal flowmetry (HRF) images when pixel‐by‐pixel analysis is performed. Methods: Thirty‐two of 732 HRF images were randomly selected from a population‐based study. Pixel‐by‐pixel analysis was performed by two trained masked graders in the following way: a square window of 40 × 40 pixels or two windows of 30 × 30 pixels or four windows of 20 × 20 pixels free from blood vessels at the peripapillary retina were identified. Using a 1 × 1‐pixel window, the grader performed pointwise analysis according to a specific protocol. The analysis process was performed by each observer three times (A, B, C) at 1‐week intervals. The percentage of pixels with < 1 arbitrary unit of flow (zero flow), the 10th, 25th, 50th, 75th and 90th percentiles and mean flow values were calculated. The difference between the results of analyses B−A and C−A for all HRF parameters was estimated using the Wilcoxon signed rank test. Mixed‐effect regression models were also used after controlling for grader effect and correlation within subjects. Results: There was no statistically significant difference between the results of analyses B−A and C−A or for any parameter in the mixed‐effect regression models. Intraclass correlation was 0.9665 for the percentage of zero flow pixels. Conclusions: Pixel‐by‐pixel analysis of HRF images by trained graders remains a highly reproducible method. No grader effect was found. If a precise protocol is followed, the results are independent of the exact placement of the analysis windows and the pointwise analysis of the identified and mapped retinal tissue.