Title: Vessel segmentation Author: Ján Dupej Department / Institute: Department of Software and Computer Science Education Supervisor of the master thesis: RNDr. Josef Pelikán, KSVI Abstract: In this thesis we researched some of the blood vessed segmentation and visualization techniques currently available for angiography on CT data. We then designed, implemented and tested a system that allows both semi-automatic and automatic vessel segmentation and visualization. For vessel segmantation and tracking we used a region-growing algorithm that we overhauled with several heuristics and combined with centerline detection. We then automated this algorithm by automatic seed generation. The visualization part is accomplished with an adaptation of the well-known straightened CPR method that we enhanced so that it visualizes the whole cross-section of the blood vessel, instead of just one line of it. Furthermore, we used the Bishop frame to maintain minimal twist of the curve-local coordinate system along the whole vessel. Keywords: vessel segmentation, medical data analysis, volume data

The image segmentation task divides an image into regions of similar pixels based on brightness, color, and texture, in which every pixel in the image is as- signed to a label. Segmentation is vital in numerous medical imaging applications, such as quantifying the size of tissues, the localization of diseases, treatment plan- ning, and surgery guidance. This thesis focuses on two medical image segmentation tasks: retinal vessel segmentation in fundus images and brain segmentation in 3D MRI images. Finally, we introduce LEON, a lightweight neural network for edge detection. The first part of this thesis proposes a lightweight neural network for retinal blood vessel segmentation. Our model achieves cutting-edge outcomes with fewer parameters. We obtained the most outstanding performance results on CHASEDB1 and DRIVE datasets with an F1 measure of 0.8351 and 0.8242, respectively. Our model has few parameters (0.34 million) compared to other networks such as ladder net with 1.5 million parameters and DCU-net with 1 million parameters. The second part of this thesis investigates the association between whole and re- gional volumetric alterations with increasing age in a large group of healthy subjects (n=6739, age range: 30–80). We used a deep learning model for brain segmentation for volumetric analysis to extract quantified whole and regional brain volumes in 95 classes. Segmentation methods are called edge or boundary-based methods based on finding abrupt changes and discontinuities in the intensity value. The third part of the thesis introduces a new Lightweight Edge Detection Network (LEON). The proposed approach is designed to integrate the advantages of the deformable unit and DepthWise Separable convolutions architecture to create a lightweight back- bone employed for efficient feature extraction. Our experiments on BSDS500 and NYUDv2 show that LEON, while requiring only 500000 parameters, outperforms the current lightweight edge detectors without using pre-trained weights. / Graduate / 2022-10-12

Knowledge of the microcirculatory system has added significant value to the analysis of tissue oxygenation and perfusion. While developments in videomicroscopy technology have enabled medical researchers and physicians to observe the microvascular system, the available software tools are limited in their capabilities to determine quantitative features of microcirculation, either automatically or accurately. In particular, microvessel density has been a critical diagnostic measure in evaluating disease progression and a prognostic indicator in various clinical conditions. As a result, automated analysis of the microcirculatory system can be substantially beneficial in various real-time and off-line therapeutic medical applications, such as optimization of resuscitation. This study focuses on the development of an algorithm to automatically segment microvessels, calculate the density of capillaries in microcirculatory videos, and determine the distribution of blood circulation. The proposed technique is divided into four major steps: video stabilization, video enhancement, segmentation and post-processing. The stabilization step estimates motion and corrects for the motion artifacts using an appropriate motion model. Video enhancement improves the visual quality of video frames through preprocessing, vessel enhancement and edge enhancement. The resulting frames are combined through an adjusted weighted median filter and the resulting frame is then thresholded using an entropic thresholding technique. Finally, a region growing technique is utilized to correct for the discontinuity of blood vessels. Using the final binary results, the most commonly used measure for the assessment of microcirculation, i.e. Functional Capillary Density (FCD), is calculated. The designed technique is applied to video recordings of healthy and diseased human and animal samples obtained by MicroScan device based on Sidestream Dark Field (SDF) imaging modality. To validate the final results, the calculated FCD results are compared with the results obtained by blind detailed inspection of three medical experts, who have used AVA (Automated Vascular Analysis) semi-automated microcirculation analysis software. Since there is neither a fully automated accurate microcirculation analysis program, nor a publicly available annotated database of microcirculation videos, the results acquired by the experts are considered the gold standard. Bland-Altman plots show that there is ``Good Agreement" between the results of the algorithm and that of gold standard. In summary, the main objective of this study is to eliminate the need for human interaction to edit/ correct results, to improve the accuracy of stabilization and segmentation, and to reduce the overall computation time. The proposed methodology impacts the field of computer science through development of image processing techniques to discover the knowledge in grayscale video frames. The broad impact of this work is to assist physicians, medical researchers and caregivers in making diagnostic and therapeutic decisions for microcirculatory abnormalities and in studying of the human microcirculation.

Image-guided therapies have the potential to improve the accuracy of treating liver cancer. In order to register intraoperative with preoperative liver images, joint segmentation and registration methods require fast segmentation of matching vessel centerlines. The algorithm presented in this thesis solves this problem by tracking the centerlines using ridge and cross-section information, and uses knowledge of the patient’s vasculature in the preoperative image to ensure correspondence. The algorithm was tested on three MR images of healthy volunteers and one CT image of a patient with liver cancer. Results show that in the context of join segmentation registration, if the registration error is less than 2.0mm, the average segmentation error is 0.73-1.68mm, with 88-100% of the vessels having an error less than a voxel length. For registration error less than 4.6mm, the average segmentation error is 1.17-2.11mm, with 79-98% of the vessels having an error less than a voxel length.

Image-guided therapies have the potential to improve the accuracy of treating liver cancer. In order to register intraoperative with preoperative liver images, joint segmentation and registration methods require fast segmentation of matching vessel centerlines. The algorithm presented in this thesis solves this problem by tracking the centerlines using ridge and cross-section information, and uses knowledge of the patient’s vasculature in the preoperative image to ensure correspondence. The algorithm was tested on three MR images of healthy volunteers and one CT image of a patient with liver cancer. Results show that in the context of join segmentation registration, if the registration error is less than 2.0mm, the average segmentation error is 0.73-1.68mm, with 88-100% of the vessels having an error less than a voxel length. For registration error less than 4.6mm, the average segmentation error is 1.17-2.11mm, with 79-98% of the vessels having an error less than a voxel length.

Indiana University-Purdue University Indianapolis (IUPUI) / Collateral connections in the brain, also known as Leptomeningeal Anastomoses, are connections between blood vessels originating from different arteries. Despite limited knowledge, they are suggested as an important contributor to cerebral stroke recovery that allows additional blood flow through the affected area. However, few databases and algorithms exist for this specific task of locating them. In this paper, a MATLAB program is developed to find these connections and detect strokes to replace manual labeling by professionals. The limited data available for this study are 23 2D microscopy images of mice cerebral vascular structures highlighted by dyes. In the images, strokes are shown to diminish the pixel count of vessels below 80\% compared to the healthy brain. Stroke classification error is greatly reduced by narrowing the scope from comparing the entire hemisphere to one smaller region. A novel way of finding collateral connections is utilizing connected components. Connected components organize all adjacent pixels into a group. All collateral connections can be found on the border of two neighboring arterial flow regions, and belong to the same group of connected components with the arterial source from each side. Along with finding collateral connections, a newly created coordinate system allows regions to be defined relative to the brain landmarks, based on the brain's center, orientation, and scale. The method newly proposed in this paper combines stroke detection, brain coordinate system extraction, and collateral connection detection in stroke-affected mouse brains using only image processing techniques. This allows a simpler, more explainable result on limited data than other techniques such as supervised machine learning. In addition, the new method does not require ground truth and high image count for training. This automated process was successfully interpreted by medical experts, which allows for further research into automating collateral connection detection in 3D.

This thesis investigates the feasibility of using fluid flow as a deformable model for segmenting vessels in 2D and 3D medical images. Exploiting fluid flow in vessel segmentation is biologically plausible since vessels naturally provide the medium for blood transportation. Fluid flow can be used as a basis for powerful vessel segmentation because streaming fluid regions can merge and split providing topological adaptivity. In addition, the fluid can also flow through small gaps formed by imaging artifacts building connections between disconnected areas. In our study, due to their simplicity, parallelism, and low computational cost compared to other fluid simulation methods, linearized shallow water equations (LSWE) are used. The method developed herein is validated using synthetic data sets, two clinical datasets, and publicly available simulated datasets which contain Magnetic Resonance Angiography (MRA) images, Magnetic Resonance Venography (MRV) images and retinal angiography images. Depending on image size, one to two order of magnitude speed ups are obtained with developed parallel implementation using Nvidia Compute Unified Device Architecture (CUDA) compared to single-core and multicore CPU implementation.

Advances in coronary Computed Tomography Angiography (CTA) have resulted in a boost in the use of this new technique in recent years, creating a challenge for radiologists due to the increasing number of exams and the large amount of data for each patient. The main goal of this study was to develop a computer tool to facilitate coronary CTA analysis by combining knowledge of medicine and image processing, and to evaluate the performance in clinical settings. Firstly, a competing fuzzy connectedness tree algorithm was developed to segment the coronary arteries and extract centerlines for each branch. The new algorithm, which is an extension of the “virtual contrast injection” (VC) method, preserves the low-density soft tissue around the artery, and thus reduces the possibility of introducing false positive stenoses during segmentation. Visually reasonable results were obtained in clinical cases. Secondly, this algorithm was implemented in open source software in which multiple visualization techniques were integrated into an intuitive user interface to facilitate user interaction and provide good over­views of the processing results. An automatic seeding method was introduced into the interactive segmentation workflow to eliminate the requirement of user initialization during post-processing. In 42 clinical cases, all main arteries and more than 85% of visible branches were identified, and testing the centerline extraction in a reference database gave results in good agreement with the gold standard. Thirdly, the diagnostic accuracy of coronary CTA using the segmented 3D data from the VC method was evaluated on 30 clinical coronary CTA datasets and compared with the conventional reading method and a different 3D reading method, region growing (RG), from a commercial software. As a reference method, catheter angiography was used. The percentage of evaluable arteries, accuracy and negative predictive value (NPV) for detecting stenosis were, respectively, 86%, 74% and 93% for the conventional method, 83%, 71% and 92% for VC, and 64%, 56% and 93% for RG. Accuracy was significantly lower for the RG method than for the other two methods (p<0.01), whereas there was no significant difference in accuracy between the VC method and the conventional method (p = 0.22). Furthermore, we developed a fast, level set-based algorithm for vessel segmentation, which is 10-20 times faster than the conventional methods without losing segmentation accuracy. It enables quantitative stenosis analysis at interactive speed. In conclusion, the presented software provides fast and automatic coron­ary artery segmentation and visualization. The NPV of using only segmented 3D data is as good as using conventional 2D viewing techniques, which suggests a potential of using them as an initial step, with access to 2D reviewing techniques for suspected lesions and cases with heavy calcification. Combining the 3D visualization of segmentation data with the clinical workflow could shorten reading time.

<p>    Collateral connections in the brain, also known as Leptomeningeal Anastomoses, are connections between blood vessels originating from different arteries. Despite limited knowledge, they are suggested as an important contributor to cerebral stroke recovery that allows additional blood flow through the affected area. However, few databases and algorithms exist for this specific task of locating them. </p> <p>    In this paper, a MATLAB program is developed to find these connections and detect strokes to replace manual labeling by professionals.  The limited data available for this study are 23 2D microscopy images of mice cerebral vascular structures highlighted by dyes. In the images, strokes are shown to diminish the pixel count of vessels below 80\% compared to the healthy brain. Stroke classification error is greatly reduced by narrowing the scope from comparing the entire hemisphere to one smaller region.</p> <p>    A novel way of finding collateral connections is utilizing connected components. Connected components organize all adjacent pixels into a group. All collateral connections can be found on the border of two neighboring arterial flow regions, and belong to the same group of connected components with the arterial source from each side. </p> <p>    Along with finding collateral connections, a newly created coordinate system allows regions to be defined relative to the brain landmarks, based on the brain's center, orientation, and scale.</p> <p>    The method newly proposed in this paper combines stroke detection, brain coordinate system extraction, and collateral connection detection in stroke-affected mouse brains using only image processing techniques.  This allows a simpler, more explainable result on limited data than other techniques such as supervised machine learning.  In addition, the new method does not require ground truth and high image count for training. This automated process was successfully interpreted by medical experts, which allows for further research into automating collateral connection detection in 3D.</p>

Retinal imaging enables the visualization of a portion of the human microvasculature in-vivo and non-invasively. The scanning laser ophthalmoscope (SLO), provides images characterized by an ultra-wide field of view (UWFoV) covering approximately 180-200º in a single scan, minimizing the discomfort for the subject. The microvasculature visible in retinal images and its changes have been vastly investigated as candidate biomarkers for different types of systemic conditions like cardiovascular disease (CVD), which currently remains the main cause of death in Europe. For the CARMEN study, UWFoV SLO images were acquired from more than 1,000 people who were recruited from two studies, TASCFORCE and SCOT-HEART, focused on CVD. This thesis presents an automated system for SLO image processing and computation of candidate biomarkers to be associated with cardiovascular risk and MRI imaging data. A vessel segmentation technique was developed by making use of a bank of multi-scale matched filters and a neural network classifier. The technique was devised to minimize errors in vessel width estimation, in order to ensure the reliability of width measures obtained from the vessel maps. After a step of refinement of the centrelines, a multi-level classification technique was deployed to label all vessel segments as arterioles or venules. The method exploited a set of pixel-level features for local classification and a novel formulation for a graph cut approach to partition consistently the retinal vasculature that was modelled as an undirected graph. Once all the vessels were labelled, a tree representation was adopted for each vessel and its branches to fully automate the process of biomarker extraction. Finally, a set of 75 retinal parameters, including information provided by the periphery of the retina, was created for each image and used for the biomarker investigation.