Data Processing for MR Spectroscopy Call for Papers

Last updated Friday, 22 August 2008

Spectral Analysis Challenge

Over the last decade there has been a tremendous increase in the both the variety and quantity of MR spectroscopy data acquired from in vivo studies. This has led to the development of new methods of analysis, beyond the traditional peak integration or single spectrum fitting procedures to extract the maximum information contained in the data. Given the wide range of possible methodologies, identifying the approach(es) which provide optimal retrieval of information for a particular type of data can be a major challenge.

The Spectral Analysis Challenge Session proposes to address this by allowing the MR spectroscopy community at large an opportunity to use their favorite analysis software on a group of datasets which have been chosen to illustrate the range of analytical problems found in modern in vivo MR spectral analysis. In this way a common group of datasets are provided which can be downloaded and analyzed by anyone interested in taking part in the challenge. Thus different approaches and software packages can be compared to one another in a rational way. Historically methods for analysis of spectroscopy data focused on quantification of the individual peaks in a single spectrum. These methods typically involve some mathematical modeling of the spectroscopy signal with the goal of obtaining numerical values for the peak-parameters (amplitude, position, width). This effort produced sophisticated and accurate approaches for quantitation of resonance peaks in a single spectrum (in either the time or frequency domain), which are widely available in the NMR community. In recent years, however, new instrumental developments: high-field magnets, fast gradients, etc., allowed the acquisition of hundreds of spectra in a short time. Chemical Shift Imaging (CSI), which can produce hundreds of spectra in a single acquisition, is now the preferred method for obtaining localized information in vivo. The datasets from kinetic in vivo experiments can also result in hundreds of spectra.

Each of these datasets contains hundreds of spectra with multiple, often overlapping peaks. The relevant information in each often differs in character. In some we seek to understand the variation in line shape in a single line. In others we need to know the temporal variation of peak intensities or perhaps the temporal variation of line shape or position. In some cases substantial prior processing is required before a final analysis step so that the overall analysis needs to be done in several passes. Efficient analysis of such datasets is quite challenging and generates a need for different methodologies than traditional single spectrum peak quantitation procedures. Although single spectrum procedures can be applied repeatedly to the individual spectra in a dataset it is difficult for them to take advantage of potential relationships among the spectra to improve the quality of the analysis. This is the potential advantage of multivariate approaches to these datasets. Nevertheless single spectrum analysis is quite sophisticated and there are clearly cases when such a method is optimal. If the Spectral Analysis Challenge can help answer which methodology is to be preferred for a given type of dataset it will have succeeded in its major purpose.

Typical preprocessing steps might be:

· Automatic routines for first-pass analysis: filtering, phasing, zerofilling, base-line correction

· Removal of artifactual variations in the data, related to shimming, field inhomogeniety, coil sensitivity, removing of
variable water-suppression artifacts, treatments of lipids, etc.

Typical information to be obtained from or questions asked of the data are:

· Co-registration of multidimensional MRS with MRIs.

· Extraction of peak intensities, either relative or absolute.

· Extraction of peak positions.

· Determination of common patterns in line shape variation.

· How many different spectral signatures are in the dataset?? Can we correlate the distribution of these signatures
with the underlying anatomical structures?

· Which features in the data best correlate with additional information about the subjects/samples under study?

· What is the best way to visualize and/or present the results?

If someone feels that they have another dataset which illustrate other novel points of spectral data analysis, please contact Truman Brown (trb11@columbia.edu) to add it to the available group. Details on how to download the datasets are given below. All results must be transmitted to Truman Brown (trb11@columbia.edu) by October 28^th, 2006 so that the session can be organized. Specific details of how to report results and what extra information will be requested for comparison purposes (type of computer, running time, etc.) will be emailed to everyone registered by October 14^th, 2006. All registrants will be asked to present a poster. Depending on their number, either summary or individual oral presentations of the various approaches and results will be made during the session.

We have assembled 4 datasets, representing variety of in vivo MRS data-types, routinely obtained from the investigators in the NMR community. We would particularly like to thank the investigators who have made these data available to make this challenge possible. To reiterate, if anyone believes that they have additional data which could be added to these, please contact Truman Brown.

The examples represent both ¹H and ³¹P NMR kinetic and spatial spectral datasets, obtained at different fields and from different anatomical structures. The datasets are:

· ^31P Human Muscle Kinetic Exercise Data (Thanks to Ronald Meyer, MSU);

· 2D ¹H water Echo Planar CSI data (HISS) from Human Breast (Thanks to Gregory Karczmar (UChicago);

· 3D ¹H Human Brain MRSI 3.0 T Spectroscopy Data (Thanks to Andrew Maudsley, UFlorida)

· 1D ¹H Human Brain 7 T kinetic data during visual stimulus (Thanks to Silvia Mangia & Ivan Tkac, CMRR, UMinn)

To register for the Spectral Analysis Challenge please contact Reza Tehrani (mt2151@columbia.edu) for a user name, password and URL location of the datasets. We are requesting registration so that we can send email to the participants. The data are maintained in separate directories with a short description. Questions should be addressed to Truman Brown (trb11@columbia.edu). Please analyze as many datasets as possible as we wish to have the best possible comparison. Thanks for taking part in this community comparison and good luck to your favorite software/analysis package! We look forward to seeing you at Airlie House in November.