Performance analysis of cone detection algorithms

Abstract

Many algorithms have been proposed to help clinicians evaluate cone density and spacing, as these may be related to the onset of retinal diseases. However, there has been no rigorous comparison of the performance of these algorithms. In addition, the performance of such algorithms is typically determined by comparison with human observers. Here we propose a technique to simulate realistic images of the cone mosaic. We use the simulated images to test the performance of three popular cone detection algorithms, and we introduce an algorithm which is used by astronomers to detect stars in astronomical images. We use Free Response Operating Characteristic (FROC) curves to evaluate and compare the performance of the four algorithms. This allows us to optimize the performance of each algorithm. We observe that performance is significantly enhanced by up-sampling the images. We investigate the effect of noise and image quality on cone mosaic parameters estimated using the different algorithms, finding that the estimated regularity is the most sensitive parameter.

Understanding the changes of cone reflectance in adaptive optics flood illumination retinal images over three years

“…Another limitation could be the automatic detection of the cones. Even if we used the algorithm with the best performance and that best suited our needs, the detection process cannot be regarded as perfect [31]. Nonetheless, we note here that the analysis of such a large amount of data would not have been possible using manual cone detection only, and one of our purposes is to move towards the automation of the analysis process.…”

“…It was carried out on all the images as well as on the average images. From previous work [31], we know that the typical mean percentage of cones correctly detected by the Chiu et al algorithm on a single image at the same eccentricities is 97%. Since the average images have more cones than any single image of the series, the algorithm performed notably better, detecting also the cones that were missed in the images because they were too faint (Fig.…”

“…Before starting the detection process, we doubled the size of the images using bicubic interpolation in order to improve the final detection performance [31]. The size on the retina of the upsampled pixels was 0.8 µm on an emmetropic eye.…”

“…It has already been observed that the position of the light source in an AO flood illumination system has an influence on the reflectance of the cone mosaic [9,24,[27][28][29], and this property has been used to study the angular tuning of single cones [15,28,30]. As the performance of the most commonly used cone detection algorithms depends on the quality of the images to which they are applied [31], it is also possible that the position of the light source itself affects not only the reflectance of the cones but also their detection and the subsequent estimation of the mosaic metrics. From a technical point of view, it would be useful for the operators of commercial AO flood illumination systems to know if there is a substantial difference in the detection results when changing the position of the light source within the central pupil area, which is the area commonly used for collecting high-quality AO images of the cone mosaic.…”

Unsupervised identification of cone photoreceptors in non-confocal adaptive optics scanning light ophthalmoscope images

“…1(b)(left)]. Several algorithms leveraged the bright appearance of the confocal images generated by AOSLO [7,8], including Circular Hough Transform [9], graph-cut with dynamicprogramming-based methods [10], and pure intensity-based methods [11].…”