Purchase this article with an account.
Robert Arnar Karlsson, Benedikt Atli Jónsson, Sveinn Hakon Hardarson, Olof Birna Olafsdottir, Gísli Hreinn Halldórsson, Einar Stefansson; Generation of features for retinal image quality evaluation using simulated annealing. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1690.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The quality of fundus images is important for successful outcome of diagnosis of various diseases.Here we present a method for automatically generating features for evaluating the image quality of retinal images from the Oxymap T1, a retinal oximeter, and compare the results with grades given by humans.
An optimization algorithm employing simulated annealing was presented with images whose qualities varied from very good to very poor. The objective of the optimization algorithm was to generate image features, by arranging simple image filters, which would allow a classifier to predict the image quality.The datasets used for training and testing was provided by Landspítali University Hospital and contained 808 retinal images captured with the Oxymap T1. The images were graded by 6 persons who were familiar with retinal images. Each grader evaluated the images separately on a scale of 0.0 to 1.0 where a higher grade indicated better quality and the grades were then averaged in order to reduce the effect of an individual grader’s bias or other human errors and gain a better estimate of an images “true” quality.A benchmark for the performance of the image evaluation algorithm was created by comparing how well the predictions of one human evaluator fit when compared with the average rating of the other evaluators. The automated algorithm was evaluated using ten-fold cross-validation where 90% of the labelled images were used to fine-tune the model parameters (training) and 10% were used for testing, this was repeated ten times with random samples being selected for testing and training.
The table below shows how well an average human grader performed versus the automated algorithm. The performance was measured using the coefficient of determination from a linear regression (R2) where the average grade of the human evaluators was considered the true quality grade.
Automatic evaluation of image quality is more accurate, faster and less subjective than analysis by a human. The automatic analysis can be used to ensure that images acquired for oximetry are of sufficient quality to be analysed.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
This PDF is available to Subscribers Only