Detection of Pathological Lesions in High Resolution Retinal Images

• Main Author: Rosander, Frida
• Format: Others
• Language: English
• Published: Linköpings universitet, Tekniska högskolan 2013
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-93837

High resolution retinal cameras with adaptive optics makes it possible to image small structures in the eye, such as photoreceptors and nerve bres, as well as blood vessels. Adaptive optics was rst developed to reduce blur in stellar images, but has later been used to correct for ocular aberrations in order to achieved higher resolution in retinal images. The development of these high resolution retinal cameras gives new pos- sibilities, and this master thesis has as purpose to investigate two of those: In-vivo estimation of cone photoreceptor distribution, and automatic detection of pathological areas in retinal images, as well as registration of retinal images. It is desirable to explore this in order to put helpful research tools into the hands of retinal researchers. For in-vivo detection of photoreceptors and calculations of cone density, the rtx1 adaptive optics retinal camera by Imagine Eyes together with the in house software AOdetect, were used. Comparison with previously published cone den- sity data showed that the in-vivo detection of photoreceptors gives an estimation of the cone densities at retinal eccentricities between 2.5 and 10 degrees. Detection of the pathological areas was performed in geographic atrophy images using an active snake contour method. It was stated that active snakes perform well, considering the diculties provided by the specic image features. However, the method has some shortcomings and it is suggested that alternative segmentation methods of atrophic areas in retinal images is further explored. In order to follow progression over time of the atrophy, an image registration algorithm has been implemented. Due to the characteristics of the geographic atrophy images, this algorithm is semi-automatic, that is, the user indicates the pairs of feature points as input. The registration performs well when the user chooses control point pairs carefully.