Umeå University's logo

Motivated by the development of optimal dispatching strategies for prehospital resources, we model the spatial distribution of ambulance call events in the Swedish municipality Skellefteå during 2014–2018 in order to identify important spatial covariates and discern hotspot regions. Our large-scale multivariate data point pattern of call events consists of spatial locations and marks containing the associated priority levels and sex labels. The covariates used are related to road network coverage, population density, and socio-economic status. For each marginal point pattern, we model the associated intensity function by means of a log-linear function of the covariates and their interaction terms, in combination with lasso-like elastic-net regularized composite/Poisson process likelihood estimation. This enables variable selection and collinearity adjustment as well as reduction of variance inflation from overfitting and bias from underfitting. To incorporate mobility adjustment, reflecting people’s movement patterns, we also include a nonparametric (kernel) intensity estimate as an additional covariate. The kernel intensity estimation performed here exploits a new heuristic bandwidth selection algorithm. We discover that hotspot regions occur along dense parts of the road network. A mean absolute error evaluation of the fitted model indicates that it is suitable for designing prehospital resource dispatching strategies. Supplementary materials accompanying this paper appear online.

In order to optimally utilise the resources of a country’s prehospital care system, i.e. ambulance service(s),it is crucial that one is able to spatio-temporally forecast hot-spots, i.e. spatial regions and periods with anincreased risk of seeing a call to the emergency number 112 which results in the dispatch of an ambulance.Such forecasts allow the dispatcher to make strategic decisions regarding e.g. the fleet size and where todirect unoccupied ambulances. In addition, simulations based on forecasts may serve as the startingpoint for different optimal routing strategies. Although the associated data typically comes in the form ofspatio-temporal point patterns, point process based modelling attempts in the literature has been scarce.In this paper, we study a unique set of Swedish spatio-temporal ambulance call data, which consists ofthe spatial (gps) locations of the dispatch addresses and the associated days of occurrence of the calls.The spatial study region is given by the four northernmost regions of Sweden and the study period isJanuary 1, 2014 to December 31, 2018. Motivated by the non-infectious disease nature of the data, wehere employ log-Gaussian Cox processes (LGCPs) for the spatio-temporal modelling and forecasting ofthe calls. To this end, we propose a K-means based bandwidth selection method for the kernel estimationof the spatial component of the separable spatio-temporal intensity function. The temporal componentof the intensity function is modelled by means of Poisson regression, using different calendar covariates,and the spatio-temporal random field component of the random intensity of the LGCP is fitted usingsimulation via the Metropolis-adjusted Langevin algorithm. A study of the spatio-temporal dynamics ofthe data shows that a hot-spot can be found in the south eastern part of the study region, where mostpeople in the region live and our fitted model/forecasts manage to capture this behaviour quite well. Thefitted temporal component of the intensity functions reveals that there is a significant association betweenthe expected number of calls and the day of the week as well as the season of the year. In addition,non-parametric second-order spatio-temporal summary statistic estimates indicate that LGCPs seem tobe reasonable models for the data. Finally, we find that the fitted forecasts generate simulated futurespatial event patterns which quite well resemble the actual future data.

The development of compressive sensing in recent years has given much attention to sparse signal recovery. In sparse signal recovery, spike and slab priors are playing a key role in inducing sparsity. The use of such priors, however, results in non-convex and mixed integer programming problems. Most of the existing algorithms to solve non-convex and mixed integer programming problems involve either simplifying assumptions, relaxations or high computational expenses. In this paper, we propose a new adaptive alternating direction method of multipliers (AADMM) algorithm to directly solve the suggested non-convex and mixed integer programming problem. The algorithm is based on the one-to-one mapping property of the support and non-zero element of the signal. At each step of the algorithm, we update the support by either adding an index to it or removing an index from it and use the alternating direction method of multipliers to recover the signal corresponding to the updated support. Moreover, as opposed to the competing “adaptive sparsity matching pursuit” and “alternating direction method of multipliers” methods our algorithm can solve non-convex problems directly. Experiments on synthetic data and real-world images demonstrated that the proposed AADMM algorithm provides superior performance and is computationally cheaper than the recently developed iterative convex refinement (ICR) and adaptive matching pursuit (AMP) algorithms.

There is a growing interest to get a fully MR based radiotherapy. The most important development needed is to obtain improved bone tissue estimation. The existing model-based methods perform poorly on bone tissues. This paper was aimed at obtaining improved bone tissue estimation. Skew-Gaussian mixture model and Gaussian mixture model were proposed to investigate CT image estimation from MR images by partitioning the data into two major tissue types. The performance of the proposed models was evaluated using the leaveone-out cross-validation method on real data. In comparison with the existing model-based approaches, the model-based partitioning approach outperformed in bone tissue estimation, especially in dense bone tissue estimation.

Two principal areas of application for estimated computed tomography (CT) images are dose calculations in magnetic resonance imaging (MRI) based radiotherapy treatment planning and attenuation correction for positron emission tomography (PET)/MRI. The main purpose of this work is to investigate the performance of hidden Markov (chain) models (HMMs) in comparison to hidden Markov random field (HMRF) models when predicting CT images of head. Obtained results suggest that HMMs deserve a further study for investigating their potential in modeling applications, where the most natural theoretical choice would be the class of HMRF models.

Purpose: There is increasing interest in computed tomography (CT) image estimations from magneticresonance (MR) images. The estimated CT images can be utilized for attenuation correction, patientpositioning, and dose planning in diagnostic and radiotherapy workflows. This study aims to introducea novel statistical learning approach for improving CT estimation from MR images and to compare theperformance of our method with the existing model-based CT image estimation methods.

Methods: The statistical learning approach proposed here consists of two stages. At the trainingstage, prior knowledge about tissue types from CT images was used together with a Gaussian mixturemodel (GMM) to explore CT image estimations from MR images. Since the prior knowledge is notavailable at the prediction stage, a classifier based on RUSBoost algorithm was trained to estimatethe tissue types from MR images. For a new patient, the trained classifier and GMMs were used topredict CT image from MR images. The classifier and GMMs were validated by using voxel-leveltenfold cross-validation and patient-level leave-one-out cross-validation, respectively.

Results: The proposed approach has outperformance in CT estimation quality in comparison withthe existing model-based methods, especially on bone tissues. Our method improved CT image estimationby 5% and 23% on the whole brain and bone tissues, respectively.

Conclusions: Evaluation of our method shows that it is a promising method to generate CTimage substitutes for the implementation of fully MR-based radiotherapy and PET/MRI applications

This thesis presents work on methods for the estimation of computed tomography (CT) images from magnetic resonance (MR) images for a number of diagnostic and therapeutic workflows. The study also demonstrates sparse signal recovery method, which is an intermediate method for magnetic resonance image reconstruction. The thesis consists of four articles. The first three articles are concerned with developing statistical methods for the estimation of CT images from MR images. We formulated spatial and non-spatial models for CT image estimation from MR images, where the spatial models include hidden Markov model (HMM) and hidden Markov random field model (HMRF) while the non-spatial models incorporate Gaussian mixture model (GMM) and skewed-Gaussian mixture model (SGMM). The statistical models are estimated via a maximum likelihood approach using the EM-algorithm in GMM and SGMM, the EM gradient algorithm in HMRF and the Baum–Welch algorithm in HMM. We have also examined CT image estimation using GMM and supervised statistical learning methods. The performance of the models is evaluated using cross-validation on real data. Comparing CT image estimation performance of the models, we have observed that GMM combined with supervised statistical learning method has the best performance, especially on bone tissues. The fourth article deals with a sparse modeling in signal recovery. Using spike and slab priors on the signal, we formulated a sparse signal recovery problem and developed an adaptive algorithm for sparse signal recovery. The developed algorithm has better performance than the recent iterative convex refinement (ICR) algorithm. The methods introduced in this work are contributions to the lattice process and signal processing literature. The results are an input for the research on replacing CT images by synthetic or pseudo-CT images, and for an efficient way of recovering sparse signal.

Denna avhandling presenterar arbete kring metoder för skattning av datortomografibilder (CT) från magnetiska resonanstomografibilder (MR) för ett antal diagnostiska och terapeutiska arbetsflöden. Studien demonstrerar även en metod för gles signalrekonstruktion, vilket är en mellanliggande metod för rekonstruktion av MR-bilder. Avhandlingen består av fyra artiklar. De tre första artiklarna handlar om att utveckla statistiska metoder för uppskattning av CT-bilder från MR-bilder. Här formuleras rumsliga och icke-rumsliga modeller för skattning av CT-bilder från MR-bilder, där de rumsliga modellerna inkluderar dolda Markov-modeller (HMM) och dolda Markov-slumpfältmodeller (HMRF), medan de icke-rumsliga modellerna består av Gaussiska mix-modeller (GMM) och skeva Gaussiska mixmodeller (SGMM). De statistiska modellerna skattas via en maximum-likelihoodansats, där EM-algoritmen används för GMM och SGMM, EM-gradientalgoritmen för HMRF samt Baum-Welch-algoritmen för HMM. Vi har även undersökt CTbildskattning med hjälp av GMM och övervakade statistiska inlärningsmetoder. Modellernas prestanda har utvärderats med hjälp av korsvalidering på faktiska data. Genom att jämföra prestandan hos modellernas CT-bildskattningar har vi observerat att GMM kombinerat med övervakad statistisk inlärning har den bästa prestandan, i synnerhet ifråga om benvävnad. Den fjärde artikeln behandlar en gles modellering inom signalrekonstruktion. Med hjälp av så kallade ”spike and slab priors” för signalen formulerade vi ett glest signalrekonstruktionsproblem och utvecklade en adaptiv algoritm för gles signalrekonstruktion. Den utvecklade algoritmen har bättre prestanda än den nyligen föreslagna iterativ konvex förfining (ICR)-algoritmen. De metoder som introducerats i detta arbete är bidrag till litteraturen inom så kallade ”lattice-processer” och signalbehandling. Resultaten levererar ett bidrag till forskningen kring ersättandet av CT-bilder med syntetiska eller pseudo-CTbilder, samt till effektiv gles signalrekonstruktion.

There is an interest to replace computed tomography (CT) images withmagnetic resonance (MR) images for a number of diagnostic and therapeuticworkflows. In this article, predicting CT images from a number of magnetic resonance imaging (MRI) sequences using regression approach isexplored. Two principal areas of application for estimated CT images aredose calculations in MRI based radiotherapy treatment planning and attenuationcorrection for positron emission tomography (PET)/MRI. Themain purpose of this work is to investigate the performance of hidden Markov (chain) models (HMMs) in comparison to hidden Markov random field (HMRF) models when predicting CT images of head. Ourstudy shows that HMMs have clear advantages over HMRF models in this particular application. Obtained results suggest that HMMs deservea further study for investigating their potential in modeling applications where the most natural theoretical choice would be the class of HMRFmodels.

There is a growing interest to get a fully MR based radiotherapy. The most important development needed is to obtain improved bone tissue estimation. Existing model-based methods have performed poorly on bone tissues. This paper aims to obtainimproved estimation of bone tissues. Skew-Gaussian mixture model (SGMM) isproposed to further investigate CT image estimation from MR images. The estimation quality of the proposed model is evaluated using leave-one-out cross-validation method on real data. In comparison with the existing model-based approaches, the approach utilized in this paper outperforms in estimation of bone tissues, especiallyon dense bone tissues.

Statistical methods are required to estimate computed tomography (CT) images from magnetic resonance (MR) images. The main purpose of estimating CT images was to get a fully MR based radiotherapy. Specifically, bone tissues and air are indistinguishable on MR images. But, there is a good contrast between soft tissue and other tissues on MR images. On CT images, there is eyecatching contrast between bone and non-bone tissues. Therefore, the main reason for CT estimation is to get improved bone tissues estimation and to use the estimated CT in fully MR based radiotherapy. The estimated CT images (also called substitute CT or Pseudo-CT images) are used for attenuation correction and dose planning in MR based radiotherapy. Gaussian mixture model (GMM) is used to investigate CT image estimation from MR images without taking spatial information into account. Markov random field (MRF) and hidden Markov model (HMM) are used to extend the approach by taking spatial dependence into account. Leave-one-dataset-out cross-validation method on five datasets (obtained from head of five patients) is used to evaluate the performance of the models. In terms of MAE, the use of spatial information improves the overall quality of CT image estimation. In this application, HMM is computationally faster and has superior performance on MRF. However, it has poor performance on bone tissues. On the other hand, MRF is computationally expensive and intractable for log-likelihood based model diagnostic. These two behaviour of HMM and MRF motivated this work to further probe the estimation of CT images from MR images by partitioning the data into bone and non-bone tissues. The partitioning of the data was based on CT value threshold. Skew-Gaussian mixture model (SGMM) and GMM applied on each partition. In terms of MAE, SGMM and GMM* (GMM applied to each partition) performed better than HMM and MRF on the bone tissues.