|
Digital image processing techniques have been widely and
successfullyapplied to many medical images. Not only is medical
image processing able to enhance critical symptoms on medical
images associated with some medical diseases for assisting
clinical diagnosis, but it can also possibly lead to a fully
automated disease diagnosis system for use by the doctors. Among
many medical images, retinal (or fundus) images provide enormous
information about diabetes, which is the leading cause of visual
impairment and blindness. The associated retinal disease is
known as diabetic retinopathy, and a specific pathology called
venous beading, provides a good indication for the degree ofthe
disease. In this thesis, we aim to develop automated diagnosis
system for venous beading. Our work consists of two main phases.
In the preprocessing phase, we develop a new method, both
effective and efficient, for extracting blood vessels of retinal
images. The method is based on the concept of Gaussian match
filters, with immediate improvement in computation time. The
extracted blood vessels are then used for diagnosing or
discriminating the disease in the second phase. The
discriminating phase adopts the approach of Shape-Cognitron
neural networks for learning and discriminating beaded veins
from normal veins. Experiment results are given and discussed,
and show great reliability and promise of the approach.
|