Diabetic retinopathy is an eye disease that causes blindness amongst many people with diabetes if left untreated. When the eye is affected, various changes in the blood vessels occur. To be able to observe these changes over time, images of the back of the eye called retinograms are acquired. Automated analysis of retinograms becomes important as the number of people afflicted with diabetes increases worldwide. A diabetic retinopathy detection system should be able to analyse retinograms, interpret them and record the changes over time. The methods associated with blood vessel detection in retinograms have a common drawback which is that small vessels that have low contrast are normally missed out during detection. Evaluation of these methods is commonly carried out using a technique which has the drawback of having a bias towards the detection of thick vessels. In this thesis a retinal vessel detection method is proposed which is capable of detecting blood vessels of different width, length and orientation in the back of the eye. The state-of-the-art method proves to be comparable to the existing methods applied to the same images used in this project. When applied for segmenting the whole vessel network, it achieves an area under the ROC curve (Az) of 0.960(±0.0021)599 compared to the best result of 0.9722 obtained via another method. When applied for vessel centreline detection, it achieves an Az of 0.977(±0.0013)66 which is higher than the best method with an Az of 0.967(±0.0017). These results are obtained using an evaluation method proposed in this project that eliminates the drawback of current evaluation methods hence removes the bias towards the detection of thick vessels.
|Date of Award||31 Dec 2014|
- The University of Manchester
|Supervisor||Chris Taylor (Supervisor) & Susan Astley (Supervisor)|
- Vessel, Centreline