Leentje Vanhamme

Brain tumor classification based on long echo proton MRS signals

There has been a growing research interest in brain tumor classification based on proton magnetic resonance spectroscopy (1H MRS) signals. Four research centers within the EU funded INTERPRET project have acquired a significant number of long echo 1H MRS signals for brain tumor classification. In this paper, we present an objective comparison of several classification techniques applied to the discrimination of four types of brain tumors: meningiomas, glioblastomas, astrocytomas grade II and metastases. Linear and non-linear classifiers are compared: linear discriminant analysis (LDA), support vector machines (SVM) and least squares SVM (LS-SVM) with a linear kernel as linear techniques and LS-SVM with a radial basis function (RBF) kernel as a non-linear technique. Kernel-based methods can perform well in processing high dimensional data. This motivates the inclusion of SVM and LS-SVM in this study. The analysis includes optimal input variable selection, (hyper-) parameter estimation, followed by performance evaluation. The classification performance is evaluated over 200 stratified random samplings of the dataset into training and test sets. Receiver operating characteristic (ROC) curve analysis measures the performance of binary classification, while for multiclass classification, we consider the accuracy as performance measure. Based on the complete magnitude spectra, automated binary classifiers are able to reach an area under the ROC curve (AUC) of more than 0.9 except for the hard case glioblastomas versus metastases. Although, based on the available long echo 1H MRS data, we did not find any statistically significant difference between the performances of LDA and the kernel-based methods, the latter have the strength that no dimensionality reduction is required to obtain such a high performance.

Frequency-domain method based on the singular value decomposition for frequency-selective NMR spectroscopy

In several applications of NMR spectroscopy the user is interested only in the components lying in a small frequency band of the spectrum. A frequency selective analysis deals precisely with this kind of NMR spectroscopy: parameter estimation of only those spectroscopic components that lie in a preselected frequency band of the NMR data spectrum, with as little interference as possible from the out-of-band components and in a computationally efficient way. In this paper we introduce a frequency-domain singular value decomposition (SVD)-based method for frequency selective spectroscopy that is computationally simple, statistically accurate, and which has a firm theoretical basis. To illustrate the good performance of the proposed method we present a number of numerical examples for both simulated and in vitro NMR data.

Does the combination of magnetic resonance imaging and spectroscopic imaging improve the classification of brain tumours

Magnetic resonance imaging (MRI) and magnetic resonance spectroscopic imaging (MRSI) play an important role in the noninvasive diagnosis of brain tumours. We investigate the use of both MRI and MRSI, separately and in combination with each other for classification of brain tissue types. Many clinically relevant classification problems are considered; for example healthy versus tumour tissues, low- versus high-grade tumours. Linear as well as nonlinear techniques are compared. The classification performance is evaluated by the area under the receiver operating characteristic (ROC) curve (AUC). In general, all techniques achieve a high performance, except when using MRI alone. For example, for low- versus high-grade tumours, low- versus high-grade gliomas, gliomas versus meningiomas, respectively a test AUC higher than 0.91, 0.93 and 0.98 is reached, when both MRI and MRSI data are used.

IQML-like algorithms for solving structured total least squares problems: a unified view

Abstract The structured total least squares (STLS) problem is an extension of the total least squares (TLS) problem for solving an overdetermined system of equations Ax ≈ b . In many cases the extended data matrix [A b] has a special structure (Hankel, Toeplitz,…). In those cases the use of STLS is often required if a maximum likelihood (ML) estimate of x is desired. The main objective of this manuscript is to clarify the difference between several IQML-like algorithms—for solving STLS problems—that have been proposed by different authors and within different frameworks. Some of these algorithms yield suboptimal solutions while others yield optimal solutions. An important result is that the classicial IQML algorithm yields suboptimal solutions to the STLS problem. We illustrate this on a specific STLS problem, namely the estimation of the parameters of superimposed exponentially damped cosines in noise. We also indicate when this suboptimality starts playing an important role.

Multichannel quantification of biomedical magnetic resonance spectroscopic signals

Quantification of individual magnetic resonance spectroscopy (MRS) signals modeled as a sum of exponentially damped sinusoids, is possible using interactive nonlinear least-squares fitting methods which provide maximum likelihood parameter estimates or using fully automatic, but statistically suboptical black-box methods. In kinetic experiments consecutive time series of MRS spectra are measured in which some of the parameters are known to remain constant over time. The purpose of this paper is to show how the previously mentioned methods can be extended to the simultaneous processing of all spectra in the time series using this additional information between the spectra. We will show that this approach yields statistically better results than processing the different signals separately.

A complete software package for MR signal processing

Magnetic resonance spectroscopy (MRS) offers a wealth of information to the biochemist or radiologist. Metabolite concentrations, J-couplings, pH, ion concentrations and gradients, temperature, etc., can all be obtained, in situ, from well-defined volumes in the human body, and in a totally non-invasive way. However, simple methods such as peak area integration or automatic line fitting in the FT MR spectrum are still relied on for routine MRS data analysis. The disadvantages of such methods are tolerated in order to keep processing fast and simple for the spectroscopist. The authors have developed a graphical user interface, in which advanced time domain signal processing methods are combined. They present a complete software package for routine MR data analysis, called MRUI, enabling the use of advanced parameter estimation algorithms with incorporation of prior knowledge via simple menus and spectral displays, in a fashion similar to the spectroscopists spectrometer software.

Java-based framework for processing and displaying short-echo-time magnetic resonance spectroscopy signals

Magnetic resonance spectroscopy (MRS) can be used to determine in a non-invasive way the concentrations of certain chemical substances, also called metabolites. The spectra of MRS signals contain peaks that correspond to the metabolites of interest. Short-echo-time signals are characterized by heavily overlapping metabolite peaks and require sophisticated processing methods. To be useful in a clinical environment tools are needed that can process those signals in an accurate and fast way. Therefore, we developed novel processing methods and we designed a freely available and open-source framework (http://www.esat.kuleuven.ac.be/sista/members/biomed) in which the processing methods can be integrated. The framework has a set of abstract classes, called hot spots, and its goal is to provide a general structure and determine the control flow of the program. It provides building blocks or components in order to help developers with integrating their methods in the framework via a plug-in system. The framework is designed with the unified modeling language (UML) and implemented in Java. When a developer implements the framework he gets an application that acts like a simple and user-friendly graphical user interface (GUI) for processing MRS data. This article describes in detail the structure and implementation of the framework and the integration of our processing methods in it.

Automatic frequency alignment and quantitation of single resonances in multiple magnetic resonance spectra via complex principal component analysis

Several algorithms for automatic frequency alignment and quantitation of single resonances in multiple magnetic resonance (MR) spectra are investigated. First, a careful comparison between the complex principal component analysis (PCA) and the Hankel total least squares-based methods for quantifying the resonances in the spectral sets of magnetic resonance spectroscopy imaging (MRSI) spectra is presented. Afterward, we discuss a method based on complex PCA plus linear regression and a method based on cross correlation of the magnitude spectra for correcting frequency shifts of resonances in sets of MR spectra. Their advantages and limitations are demonstrated on simulated MR data sets as well as on an in vivo MRSI data set of the human brain.

Magnetic resonance spectroscopic quantitation via complex principal component analysis

A careful comparison between the principal component analysis (PCA) and the Hankel total least squares (HTLS) based methods for estimating the resonances in sets of magnetic resonance (MR) spectra is presented. In addition, it is shows how to align the spectra in frequency before quantitation by applying PCA plus linear regression (LR) and cross-correlation.

Comments on “Confidence Images for MR Spectroscopic Imaging” by Karl Young, Dennis Khetselius, Brian J. Soher, and Andrew A. Maudsley (Magn Reson Med 2000;44:537–545)

The article “Confidence Images for MR Spectroscopic Imaging” by Karl Young, Dennis Khetselius, Brian J. Soher, and Andrew A. Maudsley (Magn Reson Med 2000;44:537– 545) deals with the important issue of obtaining confidence intervals for parameter estimates. However, the reasoning that leads to Eqs. [12]–[14], Figs. 2–4, and the conclusions derived from them appear to be incorrect or misstated. The authors state, “Note the perhaps nonintuitive result that Fig. 2 demonstrates, that the size of the confidence interval increases (i.e., decreased confidence and higher uncertainty) with decreasing linewidth.” The conclusions drawn by the authors reflect the statistical properties of the assessment of the height of the real part of a Lorentzian in the frequency domain, not the resonance area as stated in the text. In the following we derive the CR bounds on the amplitude and the area (related to the amplitude by a factor of 2) when the model consists of one Lorentzian line. We derive these CR bounds in two ways by carrying out the calculation in the time domain and in the frequency domain, giving of course identical results. Before starting, we define our notation: