Abstract

Abnormal tortuosity of retinal blood vessels is one of the early indicators of a number
of vascular diseases. Therefore early detection and evaluation of this phenomenon
can provide a window for early diagnosis and treatment. Currently clinicians rely on
a qualitative gross scale to estimate the degree of vessel tortuosity. There have been
many attempts to develop an accurate automated measure of tortuosity, yet it seems
that none of these measures has gained universal acceptance. This can be attributed
to the fact that descriptions and de�nitions of retinal vessel tortuosity are ambiguous
and non-standard. In addition uni�ed public datasets for di�erent disease are not
regularly available. I have propose a tortuosity evaluation framework in order to
quantify the tortuosity of arteries and veins in two dimensional colour fundus images.
The quanti�cation methods within the framework include retinal vessel morphology
analysis based on the measurements of 66 features of blood vessels. These features
are grouped as follows: 1) Structural properties 2) Distance approach features 3)
Curvature approach features 4) Combined approach features 5) Signal approach
features. The features numbered 1 to 4 above are derived from literature. Item
number �ve are new features which I have proposed and developed in this thesis.
These features have been evaluated using a manually graded retinal tortuosity
dataset as controlled set. I have also built three tortuosity datasets, each of
which contains two manual gradings. These datasets are: 1) A general tortuosity
dataset 2) A diabetic retinopathy dataset 3) A hypertensive retinopathy dataset. In
addition, I have investigated the di�erences in tortuosity patterns in hypertensive
and diabetic retinopathy. New pathology based datasets were used in this investigation.
These are the major contributions of this thesis