Wigdahl, Jeffrey
Retinal Vascular Measurement Tools for Diagnostic Feature Extraction [Tesi di dottorato]

The contributions of this work are in the development of new and state of the art algorithms for retinal image analysis including optic disc detection, tortuosity estimation, and cross-over abnormality detection. The retina is one of the only areas of the human body that blood vessels can be visualized noninvasively. Retinal imaging has become a standard in the ophthalmologist’s office because it is an easy and inexpensive way to monitor not just eye health, but also systemic vascular diseases. Changes to the retinal vasculature can be the early signs of diseases such as diabetic and hypertensive retinopathy, of which early detection can save vision, money, and improve overall health for the patient. When looking at the retinal vasculature, ophthalmologists generally rely on a qualitative assessment which can make comparisons over time or between different ophthalmologists difficult. Computer aided systems are now able to quantify what the ophthalmologist is qualitatively measuring in what they consider to be the most important features of the vasculature. These include, but are not limited to, tortuosity, arteriolar narrowing, cross-over abnormalities, and artery-vein (AV) ratio. The University of Padova has created a semi-automatic system for detecting and quantifying retinal vessels starting from optic disc detection, vessel segmentation, width estimation, tortuosity calculation, AV classification, and AV ratio. We propose a new method for optic disc detection that converts the retinal image into a graph and exploits vessel enhancement methods to calculate edge weights in finding the shortest path between pairs of points on the periphery of the image. The line segment with the maximum number of shortest paths is considered the optic disc location. The method was tested on three publicly available datasets: DRIVE, DIARETDB1, and Messidor consisting of 40, 89, and 1200 images and achieved an accuracy of 100, 98.88, and 99.42% respectively. The second contribution is a new algorithm for calculating abnormalities at AV crossing points. In retinal images, Gunn’s sign appears as a tapering of the vein at a crossing point, while Salus’s sign presents as an S-shaped curving. This work presents a method for the automatic quantification of these two signs once a crossover has been detected; combining segmentation, artery vein classification, and morphological feature extraction techniques to calculate vein widths and angles entering and exiting the crossover. Results on two datasets show separation between the two classes and that we can reliably detect and quantify these signs under the right conditions. The last contribution in tortuosity consists of two parts. A comparative study was performed on several of the most popular methods for tortuosity estimation on a new vessel dataset. Results show that several methods have good Cohen’s kappa agreement with both graders, while the tortuosity density metric has the highest single metric average agreement across vessel type and grader. The second is a new way to enhance curvature in segmented vessels based on a difference of Gabor filters to create a curvature enhanced image. The proposed method was tested on the RET-TORT database using several methods to calculate tortuosity, and had best Pearson’s correlation of .94 for arteries and .882 for veins, outperforming single mathematical formulations on the data. This held true after testing the method on the propose dataset as well, having higher correlation values across grader and vessel type compared with other tortuosity metrics. Summary of Results: The optic disc detection method was tested on three publicly available datasets: DRIVE, DIARETDB1, and Messidor consisting of 40, 89, and 1200 images and achieved an accuracy of 100, 98.88, and 99.42% respectively. The AV nicking quantification method was tested on a small dataset of 10 crossing provided by doctors at Papageorgiou Hospital, Aristotle University of Thessaloniki, Thessaloniki, Greece. Results showed separation between the normal and abnormal classes for both the Gunn and Salus sign. The method was then tested on a larger, publicly available dataset which showed good separation for the Gunn sign. The proposed tortuosity method was tested on the RET-TORT database using several methods to calculate tortuosity, and had best Pearson’s correlation of .94 for arteries and .882 for veins, outperforming single mathematical formulations on the data. It was then tested on the dataset proposed in this thesis, further corroborating the effectiveness of the method.

In relazione con http://paduaresearch.cab.unipd.it/10129/
pdf
ING-INF/06 - Bioingegneria elettronica e informatica


Tesi di dottorato. | Lingua: | Paese: | BID: TD18051915