Robust activation detection methods for real-time and offline fMRI analysis
Introduction
Magnetic resonance imaging (MRI) is a widely used technique for high resolution views of three dimensional structural images of living tissues. Functional MRI (fMRI), on the other hand, is a mode of MR imaging where the focus is on the oxygen levels in the blood, called Blood Oxygen Level Dependent (BOLD) contrast [20].
The challenge in fMRI is the constant acting of the brain, and in order to bestow a function to a particular brain region, the functional images should be captured in a carefully crafted and executed experiment. Such an experiment usually contains two alternating tasks, in which the only difference is the sought brain function. For instance, in order to locate the brain region responsible for motor functions, one task involves no physical movement, which serves as a baseline, while the alternate task involves finger tapping. Since the fMRI signals have a low signal to noise ratio (SNR) [21], [30], these tasks have to be repeated until the received signal is satisfactory. This is a block design experiment, where the stimuli and their durations are fixed, however, the methods are also expected to work on other types, such as event-related designs.
The analysis of fMRI images involves values of unit volumes called voxels. The changes in the voxel signal is inspected in order to find activated voxels in the regions responsible for the sought brain functions. In this study we propose novel approaches to identify activations in a block design fMRI experiment that can be used in either real-time or offline analysis.
The paper continues with a literature review of techniques used to estimate the activated voxels in an fMRI experiment. In Section 3.1, there is a discussion of the details of the fMRI data acquisition and experiment structure. The following section explains how the synthetic data is generated to test the proposed methods. In Section 4 the proposed methods are discussed. The results are evaluated by the receiver operating characteristic (ROC) curves and peak signal-to-noise ratio (PSNR) in Section 5. The paper concludes with the drawbacks of the proposed methods, and future work is presented.
Section snippets
Literature review
One of the earliest approaches for activation detection is the subtraction method, which is evaluated by a typical finger tapping experiment in [1]. This method simply subtracts two instances of the brain volumes related to the two tasks in order to find the activated regions. The study uses the subtraction method for the averaged volumes for each task to improve the results. Another approach in the same study is the correlation of voxel signals by the square wave form of the designed
FMRI Experiment Structure
The proposed methods are tested with the data from a two-state block design experiment acquired by the Siemens Magnetom Verio 3 Tesla device at Ege University. One of the states is called no activity, while the other is an activity, either finger tapping or word generation. The former task is related to motor regions of the brain, and the latter, to the speech center active. In the finger tapping experiment, the subjects are asked to tap their index fingers and thumbs rapidly. The tapping is
Methods
As discussed earlier, the structure of the fMRI experiment is made up of three alternating no activity and activity blocks.
The signals acquired for a voxel is denoted by the vector v(t), where t is the time a BOLD value is acquired. The acquisitions for the no activity state are denoted by the subscript vN, and acquisitions for the active state by the subscript vA. A single block of no activity or activity state is denoted by a superscript as or .
Results
For the synthetic data sets generated by Approaches 1 and 2, as discussed in Sections 3.3 and 3.4, the coordinates and the activation information of the selected voxels and their neighbors are known, making it possible to compare the activation coordinates with fully automatic activation results of the proposed methods.
The estimation performance of the proposed methods can be unveiled with the help of a binary classifier. Thus, ROC analysis is utilised as such for this study. In the light of
Discussion
Two new approaches to activation detection have been proposed in different methods. IAM, IAMP and IRRD are the instantaneous methods which introduce the application of confidence intervals on the mean of the no activity block values to decide on the activation of voxels at any activity instant. IRRD and TRRD introduce a new distance metric based on robust regression analysis that is not bound by the voxel intensities, and therefore minimizes the influence of the outliers. TAM and TRRD decide on
Acknowledgments
This work is partly funded by TUBITAK, 1001 Project Number 214S029, titled “The comparison of neural components related with default mode, short-term memory store and recall of subjects diagnosed with mild cognitive impairment with early Alzheimer’s Disease and their healthy siblings and controls.”
We extend our gratitude to the Psychiatry and Radiology Departments in the Ege University, School of Medicine, for their continuous support.
We would like to thank Simon Edward Mumford for his help in
References (34)
- et al.
Functional brain imaging at 1.5 t using conventional gradient echo mr imaging techniques
Magn. Reson. Imaging
(1993) - et al.
fMRI analysis on the GPU Possibilities and challenges
Comput. Methods Programs Biomed.
(2012) - et al.
Analysis of fMRI time series revisited
NeuroImage
(1995) - et al.
Detecting outliers in high-dimensional neuroimaging datasets with robust covariance estimators
Med. Image Anal.
(2012) Deconvolution of impulse response in event-related BOLD fmri
NeuroImage
(1999)- et al.
Variation of BOLD hemodynamic responses across subjects and brain regions and their effects on statistical analyses
NeuroImage
(2004) - et al.
The use of receiver operating characteristic curves in biomedical informatics
J. Biomed. Inf.
(2005) - et al.
Incremental activation detection in fMRI series using Kalman filtering
IEEE International Symposium on Biomedical Imaging: Nano to Macro, 2004
(2004) - et al.
Comparison of physiological noise at 1.5 t, 3 t and 7 t and optimization of fmri acquisition parameters
NeuroImage
(2005) - et al.
On the definition of signal-to-noise ratio and contrast-to-noise ratio for fmri data
PLoS ONE
(2013)
Processing strategies for time-course data sets in functional mri of the human brain
Magn. Reson. Med.
BROCCOLI: Software for fast fMRI analysis on many-core CPUs and GPUs
Front. Neuroinf.
Unsupervised learning and mapping of active brain functional mri signals based on hidden semi-markov event sequence models
Med. Imaging, IEEE Trans.
Analysis of functional MRI time-series
Hum. Brain Mapp.
Activation detection in fMRI using a maximum energy ratio statistic obtained by adaptive spatial filtering
IEEE Trans. Med. Imaging
Cited by (1)
Detection and evaluation of activation instances as change points in functional MR images
2018, 26th IEEE Signal Processing and Communications Applications Conference, SIU 2018