Purchase this article with an account.
J. F. Barraza, I. Cormensana, N. M. Grzywacz; Reading OCT Pseudo-Images: Humans vs. Bayesian Ideal Observer. Invest. Ophthalmol. Vis. Sci. 2010;51(13):1790.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Optical coherence tomography (OCT) has recently become one of the primary methods for non-invasive probing of the human retina. The pseudo-image formed by OCT (the so-called Bscan) varies probabilistically across pixels due to complexities in the measurement technique. It was shown that the distribution of pixels intensity can be well described by a stretched exponential density function such as P(I)=[1/( (β))] exp[-(I/)β ], where (β) is the gamma function, and β, and are the parameters of the distribution. In this work, we present a psychophysical experiment in which we study the ability of human subjects to detect a patch in pseudo-images and compare its performance with a Bayesian ideal observer algorithm.
The images used in this experiment were simulated and patches were created through a Gaussian modulation of parameter in the distribution. The size of the patch was defined by the standard deviation of the Gaussian. In a 4AFC experiment we measured the threshold of d necessary to discriminate the location of the patch in the image as a function of size and β, which is the parameter that controls the size of the distribution’s tail. The experiment was performed by displaying the range of intensities with both linear and logarithmic scales. Ten naïve observers took part in this experiment.
Results of the psychophysical experiment show that threshold decreases with increasing size in an exponential fashion for all βs varying from 0.6 to 1. Only for the logarithmic presentation the parameter β had an effect on the threshold. Simulations of the ideal observer show thresholds behaving in the same way as human’s data but two orders of magnitude smaller. This would show that the algorithm integrates spatial information more efficiently than humans, based on the optimal use of the knowledge it has about the stimulus.
We think this type of algorithm would be an alternative to develop systems to automatically detect anomalies in these images and then use them to help physicians during the diagnosis.
This PDF is available to Subscribers Only