Analysis of short single rest/activation epoch fMRI by self-organizing map neural network

摘要

Functional magnet resonance imaging (fMRI) has become a standard non invasive brain imaging technique delivering high spatial resolution. Brain activation is determined by magnetic susceptibility of the blood oxygen level (BOLD effect) during an activation task, e.g. motor, auditory and visual tasks. Usually box-car paradigms have 2 - 4 rest/activation epochs with at least an overall of 50 volumes per scan in the time domain. Statistical test based analysis methods need a large amount of repetitively acquired brain volumes to gain statistical power, like Student's t-test. The introduced technique based on a self-organizing neural network (SOM) makes use of the intrinsic features of the condition change between rest and activation epoch and demonstrated to differentiate between the conditions with less time points having only one rest and one activation epoch. The method reduces scan and analysis time and the probability of possible motion artifacts from the relaxation of the patients head. Functional magnet resonance imaging (fMRI) of patients for pre-surgical evaluation and volunteers were acquired with motor (hand clenching and finger tapping), sensory (ice application), auditory (phonological and semantic word recognition task) and visual paradigms (mental rotation). For imaging we used different BOLD contrast sensitive Gradient Echo Planar Imaging (GE-EPI) single-shot pulse sequences (TR 2000 and 4000, 64 $MUL 64 and 128 $MUL 128, 15 - 40 slices) on a Philips Gyroscan NT 1.5 Tesla MR imager. All paradigms were RARARA (R equals rest, A equals activation) with an epoch width of 11 time points each. We used the self-organizing neural network implementation described by T. Kohonen with a 4 $MUL 2 2D neuron map. The presented time course vectors were clustered by similar features in the 2D neuron map. Three neural networks were trained and used for labeling with the time course vectors of one, two and all three on/off epochs. The results were also compared by using a Kolmogorov-Smirnov statistical test of all 66 time points. To remove non- periodical time courses from training an auto-correlation function and bandwidth limiting Fourier filtering in combination with Gauss temporal smoothing was used. None of the trained maps, with one, two and three epochs, were significantly different which indicates that the feature space of only one on/off epoch is sufficient to differentiate between the rest and task condition. We found, that without pre-processing of the data no meaningful results can be achieved because of the huge amount of the non-activated and background voxels represents the majority of the features and is therefore learned by the SOM. Thus it is crucial to remove unnecessary capacity load of the neural network by selection of the training input, using auto-correlation function and/or Fourier spectrum analysis. However by reducing the time points to one rest and one activation epoch either strong auto- correlation or a precise periodical frequency is vanishing. Self-organizing maps can be used to separate rest and activation epochs of with only a 1/3 of the usually acquired time points. Because of the nature of the SOM technique, the pattern or feature separation, only the presence of a state change between the conditions is necessary for differentiation. Also the variance of the individual