Bukenya, Faiza
(2023)
Development of Blood Vessel Segmentation Methods for Clinical Use.
PhD thesis, University of Nottingham.
Abstract
Blood vessel segmentation is important for diagnosing disease, monitoring treatment response and for treatment planning. However, the performance of the available segmentation techniques is limited by the image artefacts (such as low contrast variation between blood vessel pixels and background, noise, and stain variation) in the medical images and therefore, the available automated segmentation techniques fail to distinguish between blood vessels from non-blood vessels. In addition, the existing segmentation techniques have a challenge of segmenting blood vessels of various sizes especially small vessels, and therefore they cannot be used during diagnosis of small blood vessel diseases (such as stroke and Alzheimer's disease) and cannot be used in medical applications that require monitoring the blood vessel growth. For a more accurate diagnosis, an automated blood vessel segmentation method should segment blood vessel of various sizes in the image. This thesis aims to develop both 3D and 2D hybrid blood vessel segmentation techniques that can overcome the challenges associated with the available methods, coupled with a high accuracy rate when applied to medical images.
Regarding the problem of image artefacts, this thesis presents a blood vessel enhancement filter for the enhancement of blood vessels in medical images referred to as a White top-hat scale-space Bilateral Hessian based Filter. The White top-hat scale-space Bilateral Hessian based Vessel Enhancement Filter (WBHVEF) is designed by integrating the capabilities of White top-hat Morphological Transform and Bilateral convolution into the Frangi Vessel Enhancement Filter to correct illumination, enhance contrast and address noise of the image while maintaining strong sharp edged blood vessels. Analysis of the blood vessel enhancement filter on the publicly available retina fundus image dataset shows that its performance is comparable to existing blood vessel enhancement filters. Moreover, the capabilities of the White top-hat scale-space Bilateral Hessian based Filter are exploited during the image enhancement stage to demonstrate its potential for improving the performance of the 3D hybrid blood vessel segmentation method and the 2D hybrid blood vessel segmentation method. Furthermore, to address the problem of lack of methods that can segment blood vessels of various sizes (especially small vessels) from 3D medical images, a 3D hybrid blood vessel segmentation method that is easily parallelizable is developed. A White top-hat scale-space Bilateral Hessian based Vessel Enhancement Filter combined with hysteresis thresholding is used during the segmentation of the small to medium blood vessels, hysteresis thresholding is used during segmentation of the medium to big vessels, image addition is utilised to combine segmentation results from the two processes, and Matlab bwareaopen operation is used to enhance the segmentation results. Evaluation of the 3D hybrid blood vessel segmentation framework on private image dataset achieved the highest average (sensitivity = 91.06\%, specificity = 99.49\%, accuracy = 99.41\%, DICE = 74.53\%, and Jaccard index = 59.40\%) on the Magnetic Resonance Angiography (MRA) carotid dataset, average (sensitivity = 89.96\%, specificity = 99.76\%, accuracy = 99.70\%, DICE = 78.93\%, and jaccard index = 65.20\%) on the Computerised Tomography (CT) pelvic arteriogram image dataset dataset, and average (sensitivity = 82.76\%, specificity = 99.62\%, accuracy = 99.47\%, DICE = 73.43\%, and jaccard index = 58.02\%) on Computed Tomography Angiography (CTA) femeral artery and veins dataset.
The capabilities of White top-hat scale-space Bilateral Hessian based Vessel Enhancement Filter, hysteresis thresholding and MATLAB bwareaopen operation are further exploited during the extraction of blood vessels from 2D retina fundus images. Evaluation of the 2D hybrid blood vessel segmentation framework on DRIVE dataset yielded state of art performance with the accuracy of 95.4\%, the sensitivity of 74.9\%, and specificity of 97.4\%. On the STARE dataset, it yielded state of art performance with the accuracy of 95.3\%, the sensitivity of 79.2\%, and specificity of 96.5\%.
In addition, this thesis presents a hybrid morphological method for the segmentation of blood vessels from the Haematoxylin and Diaminobenzidine (H\&DAB) images. The morphological method involves three steps including stain normalisation for the image enhancement, classification of pixels into blood vessel pixels and non-blood vessel pixels for segmentation and blood vessel diameter quantification for blood vessel quantification. Blood vessel diameter quantification results obtained using the proposed automated are compared to blood vessel diameter quantification results obtained using the manual method. Test results indicate that the automated is capable of achieving better results.
Regarding the issue of applicability, this thesis presents a new morphological tool for the segmentation of blood vessels from histology images to facilitate research in complex diseases such as breast cancer and Alzheimer disease.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
