Magnetic Resonance Imaging 20 (2002) 207–213

Technical note

A simple and fast technique for on-line fMRI data analysis Stefano Salvadora, Andrea Brovellib, Renata Longoa,* a

Dipartimento di Fisica, Universita’ di Trieste, Trieste, Italy b SISSA-ISAS, Cognitive Neuroscience Sector, Trieste, Italy

Received 15 December 2000; received in revised form 18 January 2002; accepted 18 January 2002

Abstract In the present work a simple technique for fMRI data analysis is presented. Artifacts due to random and stimulus-correlated motions are corrected without image registration procedures. The first step of our procedure is the calculation of the raw activation map by correlation analysis. The task related motion artifacts arise at the tissue interfaces, including vessels: when image intensity gradient is calculated the high values correspond to interface regions. To eliminate stimulus-correlated motion artifacts the intensity gradient image, obtained from the fMRI data set, is compared to the raw activation map. Since small random motions decrease the value of the correlation coefficient (R) of the external pixels of the activation areas, in the last step of our analysis procedures the clusters are extended to connected pixels having R values smaller than the defined threshold. Each cluster is expanded until the R value of the cluster average intensity is kept constant. The procedure has been tested with both GRE and EPI studies. The presented approach is a fast and robust technique useful for preliminary or on-line analysis of fMRI data. © 2002 Elsevier Science Inc. All rights reserved. Keywords: fMRI; BOLD; Data analysis

1. Introduction Extensive data manipulation is necessary to extract active brain areas from the acquired fMRI images. There are many available options for analysis and each choice may affect all subsequent results and conclusions. Therefore, fMRI data processing is an off-line procedure that takes a lot of time for both analysis procedure design and calculations [1]. The high spatial resolution (in the order of millimetres) of the fMRI images makes this technique extremely sensitive to head movements and a critical step of fMRI data analysis is the reduction of artifacts resulting from frame-to-frame motion of the subject [2]. The motion artifacts affecting fMRI examinations can be either due to stimulus-correlated movements and/or random movements. Stimulus-related movements often lead to false activated intensity time-courses in pixels at the tissue interface between high and low intensity structures (e.g., * Corresponding author. Tel.: ⫹39-040-676-3383; fax: ⫹39-040-6763350. E-mail address: [email protected] (R. Longo).

gray matter and CSF) [3]. Random movements, on the other hand, decrease the correlation between the fMRI signal and task profile in pixels on the boundary of the activation areas, thus reducing the number of pixels in the activation map. One of the most common approach for motion artifacts reduction is image realignment under rigid-body motion assumption that each image is acquired without motion artifacts. The rigid body registration hypothesis may be reasonable in EPI acquisition, but it does not fit GRE images that may suffer from motion blurring. Out-of-plane registration presents unique difficulties because data are undersampled in multislice fMRI studies, due to the slice thickness of 4 mm or larger. Moreover, since fMRI slices are not collected simultaneously, there exists no complete reference image set at any instant in time [2]. The high temporal resolution achieved with EPI technique leads to the generation of very large data sets. Iterative registration algorithms require long computing time and a non-iterative one-pass approach may be a better choice, especially in preliminary or on-line data analysis [4]. In the present work, an fMRI data analysis procedure

0730-725X/02/$ – see front matter © 2002 Elsevier Science Inc. All rights reserved. PII: S 0 7 3 0 - 7 2 5 X ( 0 2 ) 0 0 4 6 5 - 4

208

S. Salvador et al. / Magnetic Resonance Imaging 20 (2002) 207–213

based on an innovative approach to motion artifacts reduction is presented where no image registration procedures are required. The stimulus-correlated artifacts are rejected by comparison between the raw activation map and the intensity gradient image. The short computation time allows this technique to be applied to on-line or preliminary analysis of fMRI data from both GRE and EPI studies.

2. Material and Methods 2.1. Experimental design Volunteers (5 male, 4 female, aged between 26 and 35) were asked to perform self-paced single finger movements and to avoid automatic and rhythmic movements. The total acquisition time was equally divided into three motor task periods alternated with three rest periods. Seven images per each period were collected, so that a total of 42 images were acquired for each experimental session. Images were oriented transversally. The subjects’ heads were immobilized with rubber foam wedges placed between the head and the support. Field homogeneity was adjusted by means of an iterative automatic shimming procedure in each examination. The GRE studies were performed with a standard 1.5 T imager (Gyroscan S15-ACS II, Philips Medical System, Best, The Netherlands). The maximum gradient strength was 10 mT/m. A standard quadrature birdcage headcoil was used as the receiver coil, and the body coil was used for excitation. The major parameters of the 2D gradient-echo MR pulse sequence were the following: TR ⫽ 60 ms, TE ⫽ 40 ms, flip angle ⫽ 25o, FOV ⫽ 160 ⫻ 144 mm2, slice thickness ⫽ 4 mm, scan matrix ⫽ 128 ⫻ 128. The T1 contrast enhancement option was activated [5]. An MR angiography was acquired per each GRE fMRI study. The major parameters of the angiographic sequence were the following: TR ⫽ 17 ms, TE ⫽ 11 ms, flip angle ⫽ 20o, FOV ⫽ the same of the fMRI images, slice thickness ⫽ 1 mm, scan matrix ⫽ 256 ⫻ 256, slices ⫽ 12, slice thickness ⫽ 1 mm, phase contrast technique. The EPI studies were performed with use of a 1.5 T imager (Philips Gyroscan NT, Philips Medical System, Best, The Netherlands) and a standard quadrature head coil. The maximum gradient strength was 23 mT/m, the slewrate was 105 mT/m/ms. The major parameters of the 2D EPI MR pulse sequence were the following: TR ⫽ 3000 ms, TE ⫽ 45 ms, flip angle ⫽ 90o, slice thickness ⫽ 4 mm, FOV ⫽ 250 ⫻ 250 mm2, scan matrix ⫽ 64 ⫻ 128, slices ⫽ 12. The fMRI studies of a group of 8 patients (3 female, 5 man, age range 27– 60) which underwent fMRI examination with stereotaxis device before surgery [6] were reevaluated. Examinations were performed with the GRE protocol. Pa-

tients’ head was absolutely fixed in the headcoil due to stereotaxis device. This group was considered in the present study as the control group, free of motion artifacts. 2.2. Analysis methods Image analysis is performed using a program developed in IDL environment (Interactive Data Language, Research System Inc, USA). The basic analysis consists of the calculation of the correlation coefficient between time-intensity behavior of each pixel and a square wave model function. To exclude transient hemodynamic responses, 5 images per block (from the 3th to the 7th of each block) are included in the analysis. In order to reduce calculation time, the brain pixels area was extracted from the images using a simple procedure based on a morphologic filter [7]. By applying a correlation analysis (p ⬍ 0.001) and a cluster filtering (at least 5 pixels per cluster) a raw activation map has been obtained. To eliminate the stimulus-correlated artifacts, the modulus of the intensity gradient image is calculated and it is compared with the raw activation map. The intensity gradient image is obtained from the average image of the whole GRE or EPI data set. The modulus of the intensity gradient image is defined by

冘 1

G ij ⫽

兩I i, j ⫺ I i⫹n, j⫹k兩

(1)

n,k⫽⫺1

where Ii,j is the intensity of pixel (i,j) of the MR image. If the pixel is not a square, but its sides have different dimensions, the formula have to weight properly the contributions in different directions.

冘 1

G ij ⫽

P n,k兩I i, j ⫺ I i⫹n, j⫹k兩

n,k⫽⫺1

冘

(2)

1

P n,k

n,k⫽⫺1

where the Pn,k is proportional to the inverse of the distance between the center of the pixels involved. In case of EPI study using a multislice approach, the modulus of the intensity gradient image is calculated in 3D, as follows

冘 1

G ijk ⫽

P n,m,p兩I i, j,k ⫺ I i⫹n, j⫹m,k⫹p兩

n,m,p⫽⫺1

冘

(3)

1

P n,m,p

n,m,p⫽⫺1

Fig. 1 shows the modulus of the intensity gradient image of a GRE data set. The high values correspond to tissue interfaces between high and low intensity structures where the stimulus-correlated artifacts are stronger.

S. Salvador et al. / Magnetic Resonance Imaging 20 (2002) 207–213

209

Fig. 1. The modulus of intensity gradient image obtained from the average image of a GRE data set.

From the intensity gradient image a binary map is calculated by setting a threshold on the modulus value. The modulus of the intensity gradient image is normalized with respect to the maximum of the intensity histogram of the average image. The normalized images are no more affected by the experimental set-up (e.g., stereotaxis device in the field of view) and the same threshold can be applied to all studies. The threshold value has been obtained by comparing the distributions of the modulus of intensity gradient values versus the R coefficient values in health subjects and in patients. Patients’ heads were absolutely fixed in the headcoil due to stereotaxis device. In Fig. 2, the plots obtained from patients (Fig. 2a) and healthy volunteers (Fig. 2b) are presented. In stereotaxis patient studies, the pixels with significant correlation are mainly associated to low values of the intensity gradient pixels, while in volunteers experiments there are significant correlation values associated to high gradient values that are located at tissue interfaces, where stimulus-correlated motion artifacts are stronger. From the analysis of the plots in Fig. 2, we set the threshold value of the modulus of the intensity gradient to 0.16 so to calculate the binary map. To compare the raw activation map and the gradient map, we used an analysis based on clusters comparison, which is more accurate and more robust than the simple map subtraction. In both maps, each cluster has been separated and numbered using an algorithm from Serra [7]. A cluster of the raw activation map is rejected when more then 50% of its pixels are contained in a cluster of the gradient map, because this activation might be due to stimulus-correlated movements. The threshold value for the gradient map calculation has been applied to a large number of studies, in

Fig. 2. The modulus of intensity gradient image versus the correlation coefficient in 8 stereotaxic patients (2a) and in 9 volunteers (2b). (Only pixels with R ⬎ 0.45 are plotted).

order to verify that the selected value does not significantly affect the activation maps of immobilized patients, while reducing the motion artifacts in the studies performed on volunteers.

Fig. 3. The correlation coefficient of an activation cluster plotted versus the threshold value for the single pixel correlation factor.

210

S. Salvador et al. / Magnetic Resonance Imaging 20 (2002) 207–213

Fig. 4. The analysis of a GRE study: (a) the raw activation map; (b) the modulus of the intensity 2D gradient image (see text); (c) the angiographic image (see text); (d) the final activation map: light grey activation map is obtained after motion and flow artifacts rejection, the dark grey pixels are added to the clusters by the expansion procedure (see text).

In GRE study, a similar analysis is performed to reject cluster related to flow artifacts. A binary map obtained from angiographic images is used to reject clusters of the activation map in which more then 30% of the pixels are superimposed. The small random motions lower the level of correlation of the pixels on the boundary of activation areas. In order to compensate for this effect, we extended the activation cluster to the pixels, connected with the cluster, having R values smaller than the defined threshold. Before the expansion procedure, the correlation between the average cluster sig-

nal and the square–wave model signal are calculated for each activation area: the cluster is expanded until its correlation coefficient is higher than the 98% of the R values of the original cluster. After expansion the procedure of comparison with angiographic and gradient maps is repeated. In Fig. 3 the correlation coefficient of the average intensity of an activation cluster is plotted versus the threshold for the single pixel correlation coefficient. The cluster correlation coefficient is approximately constant at a highly significant value even if the R coefficient on the single pixel is decreased to about 0.3.

S. Salvador et al. / Magnetic Resonance Imaging 20 (2002) 207–213

211

Fig. 5. An EPI multislice study: light grey activation maps are obtained after motion and flow artifacts rejection, the dark grey pixels are added to the clusters by the expansion procedure (see text).

3. Results Our analysis procedure has been applied to both GRE and EPI fMRI studies. In Fig. 4, each step of the analysis

of a GRE study is presented: the raw activation map obtained from correlation analysis (Fig. 4a), the modulus of the intensity 2D gradient image (Fig. 4b), the angiographic image (Fig. 4c) and the final map (Fig. 4d). In the

212

S. Salvador et al. / Magnetic Resonance Imaging 20 (2002) 207–213

Fig. 6. Analysis of an EPI study heavily affected by motion artifacts: (a) the raw activation map; (b) the activation map after motion artifacts rejection; (c) the activation map after cluster expansion.

intensity gradient image (Fig. 4b) and in the MR angiogram (Fig. 4c) the lines define the binary maps used for the rejection of clusters possibly related to stimuluscorrelated motion artifacts and to vessels contributions. The activation maps are superimposed on a T1 weighted image, where the light grey activation map is obtained before the cluster expansion while the dark grey pixels are added to the original cluster after the expansion procedure. The slice of the angiographic image is thicker than the image slice, because vessels contained in the contiguous slices can affect the slice of interest. In Fig. 5, an EPI multislice study is shown: the light grey activation map is obtained after stimulus-correlated motion and flow artifacts rejection, the dark grey pixels are added to

the clusters by the expansion procedure. The presence of a dark grey cluster, without a light grey “seed” is due to the 3D analysis of the data: in the expansion procedure, the cluster is expanded in the neighboring slices too. Calculation time for a single-slice GRE experiment (42 acquired images) is about 20 s, and for our 12 slice EPI study (42 images per slice) the calculation time is about 2 min, using a Digital Alpha Server 4100 (Digital Unix V4.0F). The user-friendly interactive version of the program analyses a GRE single slice data set in less than 4 min. To test the degree of robustness of our technique, we performed the same analysis on an EPI study heavily affected by motion artifacts. Fig. 6 shows the raw activation

S. Salvador et al. / Magnetic Resonance Imaging 20 (2002) 207–213

213

map (Fig. 6a), the activation map after the rejection of the stimulus-correlated motion artifacts (Fig. 6b) and after cluster expansion (Fig. 6c).

In conclusion, the presented approach is a fast and robust technique useful for on-line or preliminary evaluation of the quality of the fMRI data set.

4. Discussion and conclusions

Acknowledgments

Our analysis approach does not correct the image set for the motions, but introduce some corrections on the activation map in order to reject cluster related with stimuluscorrelated motion artifacts and to compensate for low R values due to random motion at the pixels on the cluster boundary. The computing time of our procedure is very low also because it is based on the extraction of binary maps from both the correlation analysis and the intensity gradient image and on the comparison of the clusters of these maps. Operations between binary maps are obviously faster than operation between images with 12 or 16 bit per pixels. In order to reduce calculation time a new optimized version of the program can be written using C/C⫹⫹ language. The analysis procedure require the choice of a number of parameters: the lowest R values, the smallest cluster accepted, the fraction of overlapping accepted between activation cluster and gradient or angiographic maps. In the present paper, the parameters are selected in a very conservative way, in order to obtain a robust procedure for fast on-line analysis. Other fMRI data analysis software are designed for a full statistical data analysis [1], in order to obtain results in the Talairach space, after accurate image registration. They are very important tools for discussion of significant results, but they are time consuming and are not designed for a fast preliminary analysis. In order to validate our analysis procedure, our EPI data sets have been analyzed by SPM99 package [8,9]: the results are very close to the maps obtained using our technique, even if the direct comparison is not possible due to the fact that SPM results are presented in the Talairach space.

The authors kindly thank Prof. Roberto Pozzi-Mucelli and Dr. Maja Ukmar, from the Institute of Radiology of the University of Trieste, and Prof. Natale Villari, from the Institute of Radiodiagnostics of University of Florence, for their support during the experimental studies. References [1] Gold S, Christian B, Arndt S, Zeien G, Cizadlo T, Johnson DL, Flaum M, Andreasen NC. Functional MRI statistical packages: a comparative analysis. Hum Brain Mapping 1998;6:73– 84. [2] Thacker NA, Burton E, Lacey AJ, Jackson A. The effects of motion on parametric fMRI analysis techniques. Physiol Meas 1999;20:251– 63. [3] Hajnal JV, Myers R, Oatridge A, Schwieso JE, Young IR, Hydder GM. Artifacts due to stimulus correlated motion in functional imaging of the brain. Magn Reson Med 1994;31:283–91. [4] Maas LC, Frederick BdeB, Renshaw PF. Decoupled automated rotational and trasnational registration for functional MRI tima data series: the DART ragistration algorithm. Magn Reson Med 1997;37: 131–9. [5] Sobol WT, Gauntt DM. On the stationary states in gradient echo imaging. J Magn Reson Imag 1996;6:384 –98. [6] Ukmar M, Zanier F, Longo R, Rossi M, Skrap M, Pozzi-Mucelli RS. Risonanza magnetica funzionale dell’encefalo in pazienti candidati ad intervento chirurgico con guida stereotassica. Rivista di Neuroradiologia 1999;12:259 – 67. [7] Serra J. Image analysis and mathematical morphology. London: Academic Press, 1982. [8] Friston KJ, Ashburner J, Frith CD, Poline J-B, Heather JD, Frackowiak RSJ. Spatial registration and normalization of images. Hum Brain Mapping 1995;2:165– 89. [9] Friston KJ, Holmes AP, Worsley KJ, Poline J-B, Frith CD, Frackowiak RSJ. Statistical parametric maps in functional imaging: a general linear approach. Hum Brain Mapping 1995;2:189 –210.