| 1. |
- Borga, Magnus, 1965-
(författare)
-
Learning Multidimensional Signal Processing
- 1998
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- The subject of this dissertation is to show how learning can be used for multidimensional signal processing, in particular computer vision. Learning is a wide concept, but it can generally be defined as a system’s change of behaviour in order to improve its performance in some sense.Learning systems can be divided into three classes: supervised learning, reinforcement learning and unsupervised learning. Supervised learning requires a set of training data with correct answers and can be seen as a kind of function approximation. A reinforcement learning system does not require a set of answers. It learns by maximizing a scalar feedback signal indicating the system’s performance. Unsupervised learning can be seen as a way of finding a good representation of the input signals according to a given criterion.In learning and signal processing, the choice of signal representation is a central issue. For high-dimensional signals, dimensionality reduction is often necessary. It is then important not to discard useful information. For this reason, learning methods based on maximizing mutual information are particularly interesting.A properly chosen data representation allows local linear models to be used in learning systems. Such models have the advantage of having a small number of parameters and can for this reason be estimated by using relatively few samples. An interesting method that can be used to estimate local linear models is canonical correlation analysis (CCA). CCA is strongly related to mutual information. The relation between CCA and three other linear methods is discussed. These methods are principal component analysis (PCA), partial least squares (PLS) and multivariate linear regression (MLR). An iterative method for CCA, PCA, PLS and MLR, in particular low-rank versions of these methods, is presented.A novel method for learning filters for multidimensional signal processing using CCA is presented. By showing the system signals in pairs, the filters can be adapted to detect certain features and to be invariant to others. A new method for local orientation estimation has been developed using this principle. This method is significantly less sensitive to noise than previously used methods.Finally, a novel stereo algorithm is presented. This algorithm uses CCA and phase analysis to detect the disparity in stereo images. The algorithm adapts filters in each local neighbourhood of the image in a way which maximizes the correlation between the filtered images. The adapted filters are then analysed to find the disparity. This is done by a simple phase analysis of the scalar product of the filters. The algorithm can even handle cases where the images have different scales. The algorithm can also handle depth discontinuities and give multiple depth estimates for semi-transparent images.
|
|
| 2. |
- Karlsson, Anette, 1986-
(författare)
-
Quantitative Muscle Composition Analysis Using Magnetic Resonance Imaging
- 2020
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Changes in muscle tissue composition, e.g. decrease in volume and/or increase of fat infiltration, are related to adverse health conditions such as sarcopenia, inflammation, muscular dystrophy, and chronic pain. However, the onset and progression of disease and the effect of potential intervention effects are not fully understood, partly due to insufficient measurement tools. For advanced knowledge regarding these diseases, an accurate and precise measurement tool for detecting changes in muscle composition is needed. The tool must be able to detect both local changes on specific muscles for investigating local injuries and generalized muscle composition changes on a whole-body level. Magnetic resonance imaging is an excellent tool due to its superior soft tissue contrast but is normally not quantitative, making it challenging to produce reproducible results. Furthermore, manual analysis of the vast amount of images produced is extremely time consuming and therefore expensive. The aim of this thesis was to develop and validate a new magnetic resonance imaging method for muscle volume quantification and fat infiltration estimation that would have the potential to be used in both large-scale studies and for analyzing small individual muscles.The method development was divided into four main steps: 1) Rapid acquisition and reconstruction of data with sufficient resolution and calibration giving quantitative images where the relative fat content of each voxel (related to pure fat voxels) is attainable; 2) Automated muscle tissue classification based on non-rigid multi-atlas segmentation followed by probability voting to acquire the region of interest for each muscle; 3) Quantification of muscle tissue volume and fat infiltration from the classification step and the local fat signal; 4) Evaluation of the potential of the method in clinical studies.In Paper I, a method for automatic muscle volume quantification of both whole-body and regional muscles, i.e. involving steps 1–3, is presented. The automated method showed good agreement compared to manual segmentation. It was robust to an 8-fold resolution difference using two different scanner field strengths. Papers II and III evaluated the clinical relevance and the need for developing methods with high-resolution images to answer the research questions regarding the effect of a whiplash trauma on the multifidus muscles. This involved steps 1–4. The method enabled acquisition of high-resolution data to distinguish the small multifidus muscles (Paper II). The paper also showed a higher fat infiltration in the multifidus muscles in individuals with severe self-reported disability compared to individuals with milder symptoms and to healthy controls. Furthermore, the local fat infiltration was also related to widespread muscle fat infiltration (Paper III). However, the difference in widespread muscle fat infiltration could not alone distinguish between the three different groups. Paper IV showed the robustness of fat infiltration estimation when changing flip angle, and thereby the T1 weighting, of the acquired images (steps 1–3). The higher flip angle also provided better noise characteristics. Therefore, this quantitative method can be used with higher flip angle, and thus a potentially better anatomical contrast, without losing accuracy or precision.To conclude, this thesis presents a method that quantifies muscle volume and estimates fat infiltration robustly and reproducibly. The versatility of the method allows for both high-resolution images of small muscles and rapid acquisition of whole-body data. The method can be a useful tool in clinical studies regarding small individual muscles. Furthermore, the combination of being quantitative and automatic means that the method has potential to be used in longitudinal, multi-center, and large-scale studies for advanced understanding of muscular diseases.
|
|
| 3. |
- Läthén, Gunnar, 1981-
(författare)
-
Level Set Segmentation and Volume Visualization of Vascular Trees
- 2013
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Medical imaging is an important part of the clinical workflow. With the increasing amount and complexity of image data comes the need for automatic (or semi-automatic) analysis methods which aid the physician in the exploration of the data. One specific imaging technique is angiography, in which the blood vessels are imaged using an injected contrast agent which increases the contrast between blood and surrounding tissue. In these images, the blood vessels can be viewed as tubular structures with varying diameters. Deviations from this structure are signs of disease, such as stenoses introducing reduced blood flow, or aneurysms with a risk of rupture. This thesis focuses on segmentation and visualization of blood vessels, consituting the vascular tree, in angiography images.Segmentation is the problem of partitioning an image into separate regions. There is no general segmentation method which achieves good results for all possible applications. Instead, algorithms use prior knowledge and data models adapted to the problem at hand for good performance. We study blood vessel segmentation based on a two-step approach. First, we model the vessels as a collection of linear structures which are detected using multi-scale filtering techniques. Second, we develop machine-learning based level set segmentation methods to separate the vessels from the background, based on the output of the filtering.In many applications the three-dimensional structure of the vascular tree has to be presented to a radiologist or a member of the medical staff. For this, a visualization technique such as direct volume rendering is often used. In the case of computed tomography angiography one has to take into account that the image depends on both the geometrical structure of the vascular tree and the varying concentration of the injected contrast agent. The visualization should have an easy to understand interpretation for the user, to make diagnostical interpretations reliable. The mapping from the image data to the visualization should therefore closely follow routines that are commonly used by the radiologist. We developed an automatic method which adapts the visualization locally to the contrast agent, revealing a larger portion of the vascular tree while minimizing the manual intervention required from the radiologist. The effectiveness of this method is evaluated in a user study involving radiologists as domain experts.
|
|
| 4. |
- Romu, Thobias, 1984-
(författare)
-
Fat-Referenced MRI : Quantitative MRI for Tissue Characterization and Volume Measurement
- 2018
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- The amount and distribution of adipose and lean tissues has been shown to be predictive of mortality and morbidity in metabolic disease. Traditionally these risks are assessed by anthropometric measurements based on weight, length, girths or the body mass index (BMI). These measurements are predictive of risks on a population level, where a too low or a too high BMI indicates an increased risk of both mortality and morbidity. However, today a large part of the world’s population belongs to a group with an elevated risk according to BMI, many of which will live long and healthy lives. Thus, better instruments are needed to properly direct health-care resources to those who need it the most.Medical imaging method can go beyond anthropometrics. Tomographic modalities, such as magnetic resonance imaging (MRI), can measure how we have stored fat in and around organs. These measurements can eventually lead to better individual risk predictions. For instance, a tendency to store fat as visceral adipose tissue (VAT) is associated with an increased risk of diabetes type 2, cardio-vascular disease, liver disease and certain types of cancer. Furthermore, liver fat is associated with liver disease, diabetes type 2. Brown adipose tissue (BAT), is another emerging component of body-composition analysis. While the normal white adipose tissue stores fat, BAT burns energy to produce heat. This unique property makes BAT highly interesting, from a metabolic point of view.Magnetic resonance imaging can both accurately and safely measure internal adipose tissue compartments, and the fat infiltration of organs. Which is why MRI is often considered the reference method for non-invasive body-composition analysis. The two major challenges of MRI based body-composition analysis are, the between-scanner reproducibility and a cost-effective analysis of the images. This thesis presents a complete implementation of fat-referenced MRI, a technique that produces quantitative images that can increase both inter-scanner and automation of the image analysis.With MRI, it is possible to construct images where water and fat are separated into paired images. In these images, it easy to depict adipose tissue and lean tissue structures. This thesis takes water-fat MRI one step further, by introducing a quantitative framework called fat-referenced MRI. By calibrating the image using the subjects' own adipose tissue (paper II), the otherwise non-quantitative fat images are made quantitative. In these fat-referenced images it is possible to directly measure the amount of adipose tissue in different compartments. This quantitative property makes image analysis easy and accurate, as lean and adipose tissues can be separated on a sub-voxel level. Fat-referenced MRI further allows the quantification and characterization of BAT.This thesis work starts by formulating a method to produce water-fat images (paper I) based on two gradient recall images, i.e.\ 2-point Dixon images (2PD). It furthers shows that fat-referenced 2PD images can be corrected for T2*, making the 2PD body-composition measurements comparable with confounder-corrected Dixon measurements (paper III}).Both the water-fat separation method and fat image calibration are applied to BAT imaging. The methodology is first evaluated in an animal model, where it is shown that it can detect both BAT browning and volume increase following cold acclimatization (paper IV). It is then applied to postmortem imaging, were it is used to locate interscapular BAT in human infants (paper V). Subsequent analysis of biopsies, taken based on the MRI images, showed that the interscapular BAT was of a type not previously believed to exist in humans. In the last study, fat-referenced MRI is applied to BAT imaging of adults. As BAT structures are difficult to locate in many adults, the methodology was also extended with a multi-atlas segmentation methods (paper VI).In summary, this thesis shows that fat-referenced MRI is a quantitative method that can be used for body-composition analysis. It also shows that fat-referenced MRI can produce quantitative high-resolution images, a necessity for many BAT applications.
|
|
| 5. |
- Rydell, Joakim, 1979-
(författare)
-
Advanced MRI Data Processing
- 2007
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Magnetic resonance imaging (MRI) is a very versatile imaging modality which can be used to acquire several different types of images. Some examples include anatomical images, images showing local brain activation and images depicting different types of pathologies. Brain activation is detected by means of functional magnetic resonance imaging (fMRI). This is useful e.g. in planning of neurosurgical procedures and in neurological research. To find the activated regions, a sequence of images of the brain is collected while a patient or subject alters between resting and performing a task. The variations in image intensity over time are then compared to a model of the variations expected to be found in active parts of the brain. Locations with high correlation between the intensity variations and the model are considered to be activated by the task.Since the images are very noisy, spatial filtering is needed before the activation can be detected. If adaptive filtering is used, i.e. if the filter at each location is adapted to the local neighborhood, very good detection performance can be obtained. This thesis presents two methods for adaptive spatial filtering of fMRI data. One of these is a modification of a previously proposed method, which at each position maximizes the similarity between the filter response and the model. A novel feature of the presented method is rotational invariance, i.e. equal sensitivity to activated regions in different orientations. The other method is based on bilateral filtering. At each position, this method averages pixels which are located in the same type of brain tissue and have similar intensity variation over time.A method for robust correlation estimation is also presented. This method automatically detects local bursts of noise in a signal and disregards the corresponding signal segments when the correlation is estimated. Hence, the correlation estimate is not affected by the noise bursts. This method is useful not only in analysis of fMRI data, but also in other applications where correlation is used to determine the similarity between signals.Finally, a method for correcting artifacts in complex MR images is presented. Complex images are used e.g. in the Dixon technique for separate imaging of water and fat. The phase of these images is often affected by artifacts and therefore need correction before the actual water and fat images can be calculated. The presented method for phase correction is based on an image integration technique known as the inverse gradient. The method is shown to provide good results even when applied to images with severe artifacts.
|
|
| 6. |
- Sundvall, Erik, 1973-
(författare)
-
Scalability and Semantic Sustainability in Electronic Health Record Systems
- 2013
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- This work is a small contribution to the greater goal of making software systems used in healthcare more useful and sustainable. To come closer to that goal, health record data will need to be more computable and easier to exchange between systems.Interoperability refers to getting systems to work together and semantics concerns the study of meanings. If Semantic interoperability is achieved then information entered in one information system is usable in other systems and reusable for many purposes. Scalability refers to the extent to which a system can gracefully grow by adding more resources. Sustainability refers more to how to best use available limited resources. Both aspects are important.The main focus and aim of the thesis is to increase knowledge about how to support scalability and semantic sustainability. It reports explorations of how to apply aspects of the above to Electronic Health Record (EHR) systems, associated infrastructure, data structures, terminology systems, user interfaces and their mutual boundaries.Using terminology systems is one way to improve computability and comparability of data. Modern complex ontologies and terminology systems can contain hundreds of thousands of concepts that can have many kinds of relationships to multiple other concepts. This makes visualization challenging. Many visualization approaches designed to show the local neighbourhood of a single concept node do not scale well to larger sets of nodes. The interactive TermViz approach described in this thesis, is designed to aid users to navigate and comprehend the context of several nodes simultaneously. Two applications are presented where TermViz aids management of the boundary between EHR data structures and the terminology system SNOMED CT.The amount of available time from people skilled in health informatics is limited. Adequate methods and tools are required to develop, maintain and reuse health-IT solutions in a sustainable way. Multiple levels of modelling including a fixed reference model and another layer of flexible reusable ‘archetypes’ for domain specific data structures, is an approach with that aim used in openEHR and the ISO 13606 standard. This approach, including learning, implementing and managing it, is explored from different angles in this thesis. An architecture applying Representational State Transfer (REST) to archetype-based EHR systems, in order to address scalability, is presented. Combined with archetyping this architecture also aims at enabling a sustainable way of continuously evolving multi-vendor EHR solutions. An experimental open source implementation of it, aimed for learning and prototyping, is also presented.Manually changing database structures used for storage every time new versions of archetypes and associated data structures are needed is likely not a sustainable activity. Thus storage systems that can handle change with minimal manual interventions are desirable. Initial explorations of performance and scalability in such systems are also reportedGraphical user interfaces focused on EHR navigation, time-perspectives and highlighting of EHR content are also presented – illustrating what can be done with computable health record data and the presented approaches.Desirable aspects of semantic sustainability have been discussed, including: sustainable use of limited resources (such as available time of skilled people), and reduction of unnecessary risks. A semantic sustainability perspective should be inspired and informed by research in complex systems theory, and should also include striving to be highly aware of when and where technical debt is being built up. Semantic sustainability is a shared responsibility.The combined results presented contribute to increasing knowledge about ways to support scalability and semantic sustainability in the context of electronic health record systems. Supporting tools, architectures and approaches are additional contributions.
|
|
