Search

Among the methods proposed for the analysis of functional MR we have previously introduced a model-independent analysis based on the self-organizing map (SOM) neural network technique. The SOM neural network can be trained to identify the temporal patterns in voxel time-series of individual functional MRI (fMRI) experiments. The separated classes consist of activation, deactivation and baseline patterns corresponding to the task-paradigm. While the classification capability of the SOM is not only based on the distinctness of the patterns themselves but also on their frequency of occurrence in the training set, a weighting or selection of voxels of interest should be considered prior to the training of the neural network to improve pattern learning. Weighting of interesting voxels by means of autocorrelation or F-test significance levels has been used successfully, but still a large number of baseline voxels is included in the training. The purpose of this approach is to avoid the inclusion of these voxels by using three different levels of segmentation and mapping from Talairach space: (1) voxel partitions at the lobe level, (2) voxel partitions at the gyrus level and (3) voxel partitions at the cell level (Brodmann areas). The results of the SOM classification based on these mapping levels in comparison to training with all brain voxels are presented in this paper.

In the last few years more and more University Hospitals as well as private hospitals changed to digital information systems for patient record, diagnostic files and digital images. Not only that patient management becomes easier, it is also very remarkable how clinical research can profit from Picture Archiving and Communication Systems (PACS) and diagnostic databases, especially from image databases. Since images are available on the finger tip, difficulties arise when image data needs to be processed, e.g. segmented, classified or co-registered, which usually demands a lot computational power. Today's clinical environment does support PACS very well, but real image processing is still under-developed. The purpose of this paper is to introduce a parallel cluster of standard distributed systems and its software components and how such a system can be integrated into a hospital environment. To demonstrate the cluster technique we present our clinical experience with the crucial but cost-intensive motion correction of clinical routine and research functional MRI (fMRI) data, as it is processed in our Lab on a daily basis.

AnalysisofEvent-RelatedfMRIusingaNon-LinearRegressionSelf-OrganizingMapNeuralNetworkStephanG. Erb ericha, ManouLieb ert,Klaus WillmesbArmin Thronc,Walter Ob erschelpdaNeurofunctional Imaging Lab oratory,Interdisciplinary Centerfor Clinical Researchin theNeurosciences,Univ.Hospital, Univ.of TechnologyRWTH,Aachen,GermanybDept.ofNeuropsychology,Univ.Hospital, Univ.of TechnologyRWTH,Aachen,GermanycDept.of Neuroradiology,Univ.Hospital, Univ.of TechnologyRWTH,Aachen,GermanydDept.for Computer ScienceVI I,Univ.of TechnologyRWTH,Aachen,GermanyABSTRACTFunctional magnetic resonance imaging (fMRI) b ecomes a common metho d to study task induced brain activation.Using rapid Echo Planar Imaging (EPI) sequences one can obtain a higher MR-Signal under a task condition closebyactivated areas as a result of susceptibilitychanges in blo o d oxygenation (BOLD e ect).Beside the commonlyused blo cked task designs, event-related paradigms gain more imp ortance for activation of higher cognitive functionsenablingmoresophisticatedandcomplexparadigms.Fortheanalysisofevent-relatedfMRIdataonecanusestatisticaltests,inexamplet-testusedbySPMSoftware.Thetro ducedanalysismetho dbasedonanarti cialneural network algorithm, a self-organizing map(SOM),is capable to distinguish b eteen task related activation,deactivationandbaselinepatternsfromthetimeseries.Thisisachievedbytemp oralsortingpro jectionofall events from one condition into one combined hemo dynamic resp onse sampling for eachvoxel.These resp onses,having individual patterns can b e separated by their pattern features and is done by training of the neural network.After training the SOM consists of a pattern-to-voxel mapping which is sup erimp osed onto either an anatomical orEPI image of the sub ject for the task evaluation.Keywords:fMRI, non-linear regression, event-related fMRI, self-organizing map (SOM), time series analysis, un-sup ervised classi cation1.INTRODUCTIONEvent-Related functionalMRI(efMRI)hasb ecomeama jor paradigmtechniqueforBOLD(Blo o dOxygen LevelDep endent) fMRI studies.While an event-related design overcomes the lack of random stimuli presentation of b oxcar designs, one can presentstimuli without training e ects by the sub ject.Also random combinations of more thanone eventtyp e can b e investigated at once and failed evt reactions by the sub ject can b e indep endently remoedfrom the later analysis.Therefore efMRI techniques gain more and more imp ortance.A common to olb ox for the statistical analysis is the Statistical Parametric Mapping (SPM) software [1].SPM99,the latest release, uses a general linear mo del approach to describ e the time series of eachvoxel.While a linear mo delbased metho d can only explain e ects describ ed bythe chosen mo del, weinvestigated a new data-driven metho d toanalyze the fMRItime-series.This metho d is based on an self-organizing neural network, theself-organizing map[2].The neural network is trained by feature vectors, each consists of the sorted adjusted temp oral resp onse for oneeventtyp e of the time series from a voxel.This non-linear approach is based on the similarity measure, the Euclideandistance, b etween inputHRFand the trained resp onse intheneural netork.As cluster techniques [3],the SOMneeds this similarity measure to depict the b est matching neuron.Toaverage the vector p ointwecho ose the SPM99 HRF- tting function.The adjusted and tted resp onses fromeachvoxelarede ningthetrainingandlaterclassi cationinputforneuralnetwork.Afterneuron weights(vectors) representtheresp onse features,curvecharacteristics, ofpresented ttedtimeseries.*Corresp ondence: E-mail: Stephan.Erb erich@izkf.rwth-aachen.de; Phone:++49-241-80-88890; Fax: ++49-241-88-90505

The structure of an fMRI time series coregistration algorithm can be divided into modules (preprocessing, minimization procedure, interpolation method, cost function), for each of which there are many different approaches. In our study we implemented some of the most recent techniques and compared their combinations with regard to both registration accuracy and runtime performance. Bidirectional inconsistency and difference image analysis served as quality measures. The result shows that with an appropriate choice of methods realignment results can be improved by far compared with standard solutions. Finally, an automatic parameter adaptation method was incorporated. Additionally, the algorithm was implemented to run on a distributed 48 processor PC cluster, surpassing the performance of conventional applications running on high end workstations.