Digital analysis of high-resolution fundus images

Date of Award




Degree Name

Doctor of Philosophy (Ph.D.)


Biomedical Engineering

First Committee Member

Joachim H. Nagel, Committee Chair


High resolution digital color fundus images were used to investigate the tapetal-like reflex, a retinal reflection uniquely present in carriers of X-linked retinitis pigmentosa (XLRP). A Zeiss fundus camera and color transparency film were used for the fundus photography, and a slide scanner was used to digitize the color slides (6$\mu$m/pixel). A model of the imaging system was derived based on its four components (eye, camera, film, and scanner). A regularized linear restoration method based on a simplified imaging system model, was developed to reduce the image noise and recover the high frequency information. A multi-scale segmentation method was developed to separate the tapetal-like reflex patches from the retinal background. The segmentation method was based on local thresholding using two differently sized neighborhoods.In order to analyze the reflex progression over time, a registration method was developed for the case when differences between images can be approximated with translation, rotation, and scaling. The method was based on a global optimization in a log-polar transformed discrete parameter space. Validation with simulated images and cross-validation with multiple fundus images of the same region showed that an accuracy of 0.07 degrees in rotation, 0.1% in scaling, and 0.3 pixels in translation is attained.Analysis of fundus photographs of four XLRP carriers showed that the patches making up the tapetal-like reflex have, on the average, 30% higher reflectance compared to non-reflex retina. Mathematical morphology methods applied to segmented images suggested that the reflex is made of elongated structures with a small dimension of less than 14$\mu$m (on the retina) and a preferred orientation towards the fovea. In the case of one patient where fundus photographs of both eyes separated by 23 years were available, a comparison suggested that there were no detectable changes in the reflex over this time period.Segmentation and registration methods developed in this dissertation are potentially applicable to other medical and nonmedical images. The ability to reproducibly detect tiny lesions from fundus photographs should be useful for quantification of the progression of many different fundus lesions representing manifestations of a variety of systemic and ocular diseases.


Health Sciences, Ophthalmology; Engineering, Biomedical; Engineering, Electronics and Electrical

Link to Full Text