Medical Informatics

Functional magnetic resonance imaging (fMRI) is a prime example of multi-disciplinary research. Without the beautiful physics of MRI, there wouldnot be any images to look at in the first place. To obtain images of goodquality, it is necessary to fully understand the concepts of the frequencydomain. The analysis of fMRI data requires understanding of signal pro-cessing, statistics and knowledge about the anatomy and function of thehuman brain. The resulting brain activity maps are used by physicians,neurologists, psychologists and behaviourists, in order to plan surgery andto increase their understanding of how the brain works.

This thesis presents methods for real-time fMRI and non-parametric fMRIanalysis. Real-time fMRI places high demands on the signal processing,as all the calculations have to be made in real-time in complex situations.Real-time fMRI can, for example, be used for interactive brain mapping.Another possibility is to change the stimulus that is given to the subject, inreal-time, such that the brain and the computer can work together to solvea given task, yielding a brain computer interface (BCI). Non-parametricfMRI analysis, for example, concerns the problem of calculating signifi-cance thresholds and p-values for test statistics without a parametric nulldistribution.

Two BCIs are presented in this thesis. In the first BCI, the subject wasable to balance a virtual inverted pendulum by thinking of activating theleft or right hand or resting. In the second BCI, the subject in the MRscanner was able to communicate with a person outside the MR scanner,through a virtual keyboard.

A graphics processing unit (GPU) implementation of a random permuta-tion test for single subject fMRI analysis is also presented. The randompermutation test is used to calculate significance thresholds and p-values forfMRI analysis by canonical correlation analysis (CCA), and to investigatethe correctness of standard parametric approaches. The random permuta-tion test was verified by using 10 000 noise datasets and 1484 resting statefMRI datasets. The random permutation test is also used for a non-localCCA approach to fMRI analysis.

The validity of parametric functional magnetic resonance imaging (fMRI) analysis has only been reported for simulated data.Recent advances in computer science and data sharing make it possible to analyze large amounts of real fMRI data. In this study,1484 rest datasets have been analyzed in SPM8, to estimate true familywise error rates. For a familywise significance threshold of5%, significant activity was found in 1% - 70% of the 1484 rest datasets, depending on repetition time, paradigm and parametersettings. This means that parametric significance thresholds in SPM both can be conservative or very liberal. The main reason forthe high familywise error rates seems to be that the global AR(1) auto correlation correction in SPM fails to model the spectra ofthe residuals, especially for short repetition times. The findings that are reported in this study cannot be generalized to parametricfMRI analysis in general, other software packages may give different results. By using the computational power of the graphicsprocessing unit (GPU), the 1484 rest datasets were also analyzed with a random permutation test. Significant activity was thenfound in 1% - 19% of the datasets. These findings speak to the need for a better model of temporal correlations in fMRI timeseries.

Background

Procedures documented by general practitioners in primary care have not been studied in relation to procedure coding systems. We aimed to describe procedures documented by Swedish general practitioners in electronic patient records and to compare them to the Swedish Classification of Health Interventions (KVÅ) and SNOMED CT.

Methods

Procedures in 200 record entries were identified, coded, assessed in relation to two procedure coding systems and analysed.

Results

417 procedures found in the 200 electronic patient record entries were coded with 36 different Classification of Health Interventions categories and 148 different SNOMED CT concepts. 22.8% of the procedures could not be coded with any Classification of Health Interventions category and 4.3% could not be coded with any SNOMED CT concept. 206 procedure-concept/category pairs were assessed as a complete match in SNOMED CT compared to 10 in the Classification of Health Interventions.

Conclusions

Procedures documented by general practitioners were present in nearly all electronic patient record entries. Almost all procedures could be coded using SNOMED CT.

Classification of Health Interventions covered the procedures to a lesser extent and with a much lower degree of concordance. SNOMED CT is a more flexible terminology system that can be used for different purposes for procedure coding in primary care.

Functional magnetic resonance imaging (fMRI) makes it possible to non-invasively measure brain activity with high spatial resolution.There are however a number of issues that have to be addressed. One is the large amount of spatio-temporal data that needsto be processed. In addition to the statistical analysis itself, several preprocessing steps, such as slice timing correction and motioncompensation, are normally applied. The high computational power of modern graphic cards has already successfully been used forMRI and fMRI. Going beyond the first published demonstration of GPU-based analysis of fMRI data, all the preprocessing stepsand two statistical approaches, the general linear model (GLM) and canonical correlation analysis (CCA), have been implementedon a GPU. For an fMRI dataset of typical size (80 volumes with 64 x 64 x 22 voxels), all the preprocessing takes about 0.5 s on theGPU, compared to 5 s with an optimized CPU implementation and 120 s with the commonly used statistical parametric mapping(SPM) software. A random permutation test with 10 000 permutations, with smoothing in each permutation, takes about 50 s ifthree GPUs are used, compared to 0.5 - 2.5 h with an optimized CPU implementation. The presented work will save time forresearchers and clinicians in their daily work and enables the use of more advanced analysis, such as non-parametric statistics, bothfor conventional fMRI and for real-time fMRI.

Image registration is frequently used within the medical image domain and where methods with high performance are required. The need for high accuracy coupled with high speed is especially important for applications such as adaptive radiation therapy and image-guided surgery. During the last years, a number of significant projects have been introduced to make the computational power of GPUs available to a wider audience. The most well known project is the introduction of CUDA (Compute Unified Device Architecture). In this paper, we present a CUDA based GPU implementation of a non-rigid image registration algorithm, known as the Morphon, and compare it with a CPU implementation of the Morphon. The achieved speedup, in the range of 51-54x, is also compared with speedups reported from other non-rigid registration methods mplemented on the GPU. These include the Demons algorithm and a mutual information based algorithm.

The multilinear least-squares (MLLS) problem is an extension of the linear least-squares problem. The difference is that a multilinearoperator is used in place of a matrix-vector product. The MLLS istypically a large-scale problem characterized by a large number of local minimizers. It originates, for instance, from the design of filter networks. We present a global search strategy that allows formoving from one local minimizer to a better one. The efficiencyof this strategy isillustrated by results of numerical experiments performed forsome problems related to the design of filter networks.

The number of available imaging modalities increases both in clinical practice and in clinical studies. Even though data from multiple modalities might be available, image registration is typically only performed using data from a single modality. In this paper, we propose using certainty maps together with multi-channel diffeomorphic demons in order to improve both accuracy and robustness when performing image registration. The proposed method is evaluated using DTI data, multiple region overlap measures and a fiber bundle similarity metric.

Canonical correlation analysis (CCA) is a statistical methodthat can be preferable to the general linear model (GLM) for analysisof functional magnetic resonance imaging (fMRI) data. There are,however, two problems with CCA based fMRI analysis. First, it is notfeasible to use a parametric approach to calculate an activity thresholdfor a certain signi cance level. Second, two covariance matrices need tobe estimated in each voxel, from a rather small number of time samples.We recently solved the rst problem by doing random permutation testson the graphics processing unit (GPU), such that the null distribution ofany maximum test statistics can be estimated in the order of minutes. Inthis paper we consider the second problem. We extend the idea of variancepooling, that previously has been used for the GLM, to covariancepooling to improve the estimates of the covariance matrices. Our GPUimplementation of random permutation tests is used to calculate signicance thresholds, which are needed to compare the di erent activitymaps in an objective way. The covariance pooling results in more robustestimates of the covariance matrices. The number of signi cantly activevoxels that are detected (thresholded at p = 0.05, corrected for multiplecomparisons) is increased with 40 - 120% (if 8 mm smoothing is appliedto the covariance estimates). Too much covariance pooling can howeverresult in a loss of small activity clusters, 7-10 mm of smoothing givesthe best results. The calculations that were made in order to generatethe results in this paper would have taken a total of about 65 days witha Matlab implementation and about 10 days with a multithreaded Cimplementation, with our multi-GPU implementation they took about 2hours. By using fast random permutation tests, suggested improvementsof existing methods for fMRI analysis can be evaluated in an objective way.

A well-known approach to the design of computationally efficient filters is to use spectral factorization, i.e. a decomposition of a filter into a sequence of sub-filters. Due to the sparsity of the sub-filters, the typical processing speedup factor is within the range 1-10 in 2D, and for 3D it achieves10-100. The design of such decompositions consists in choosing the proper number of sub-filters, their individual types and sparsity. We focus here on finding optimal values of coefficients for given sequences of sparse sub-filters. It is a non-convex large scale optimization problem. The existing approaches are characterized by a lack of robustness - a very slow convergence with no guarantee of success. They are typically based on generating random initial points for further refinement with the use of local search methods. To deal with the multi-extremal nature of the original problem, we introduce a new constrained optimization problem. Its solution is then used as an initial point in the original problem for further refinement. Our approach is applicable to designing multi-dimensional filters. Its efficiency and robustness is illustrated by designing sub-filter sequences for 2D low-pass, band-pass and high-pass filters of approximately the same quality as with the use of a standard approach, but with the overall design speedup factor of several hundred.

Parametric statistical methods, such as Z-, t-, and F-values are traditionally employed in functional magnetic resonance imaging (fMRI) for identifying areas in the brain that are active with a certain degree of statistical significance. These parametric methods, however, have two major drawbacks. First, it is assumed that the observed data are Gaussian distributed and independent; assumptions that generally are not valid for fMRI data. Second, the statistical test distribution can be derived theoretically only for very simple linear detection statistics. With non-parametric statistical methods, the two limitations described above can be overcome. The major drawback of non-parametric methods is the computational burden with processing times ranging from hours to days, which so far have made them impractical for routine use in single subject fMRI analysis. In this work, it is shown how the computational power of cost-efficient Graphics Processing Units (GPUs) can be used to speed up random permutation tests. A test with 10 000 permutations takes less than a minute, making statistical analysis of advanced detection methods in fMRI practically feasible. To exemplify the permutation based approach, brain activity maps generated by the General Linear Model (GLM) and Canonical Correlation Analysis (CCA) are compared at the same significance level. During the development of the routines and writing of the paper, 3-4 years of processing time has been saved by using the GPU.

The need of image registration is increasing, especially in the medical image domain. The simplest kind of image registration is to match two images that have similar intensity. More advanced cases include the problem of registering images of different intensity, for which phase based algorithms have proven to be superior. In some cases the phase based registration will fail as well, for instance when the images to be registered do not only differ in intensity but also in local phase. This is the case if a dark circle in the reference image is a bright circle in the source image. While rigid registration algorithms can use other parts of the image to calculate the global transformation, this problem is harder to solve for non-rigid registration. The solution that we propose in this work is to use the local phase of the magnitude of the local structure tensor, instead of the local phase of the image intensity. By doing this, we achieve invariance both to the image intensity and to the local phase and thereby only use the structural information, i.e. the shapes of the objects, for registration.