Increased detection sensitivity in fMRI by adaptive filtering.

(1)

0 0.36 0.32 0.58 0.22 0.59 0.34 0.62 Correlation Correlation Correlation SNR Activation patterns Ordinary correlation Pre−smoothed correlation CCA x4 (t) x5 (t) x3 (t) x2 (t) x1 (t)

by Canonical Correlation Analysis.

A better approach is however offered

basis functions which correlates the most with the voxel timecourse.

t

A single timecourse y(t) is not able to model the variations in the BOLD response.

fMRI images with a Gaussian shaped filter.

It is well known that these methods are equivalent.

x(t) = w x1 x1 (t) + .... + w x5 x5 (t) ρ

A method for detecting neural activity in BOLD fMRI is

Detection sensitivity is improved with the Canonical Correlation method. Smoothing the 3x3 region

y(t) = w y1 y1 (t) + .... + w y6 y6 (t) y(t) = w y1 y1 (t) + .... + w y6 y6 (t) x(t) = w x1 x1 (t) + .... + w x5 x5 (t)

By partitioning the 3x3 region as indicated to the right, different symmetric spatial adaptive filters are obtained. Multiple _Correlation

BOLD model basis functions

Cross− Correlation . . . n 1 y y

[2] Ardekani B et al. (1998) Magn. Reson. Imag. 16(10) : 1217−1225

Example:

CONCLUSION

A widely used approach is to smooth the

CONTACT

ORDINARY CORRELATION ANALYSIS

INCREASED DETECTION SENSITIVITY IN FMRI BY ADAPTIVE FILTERING

derived as a natural extension of established analysis methods. The new method, which is based on Canonical Correlation Analysis, has an inherent adaptive spatial filtering property which improves the detection sensitivity.

A Multiple Correlation Analysis between an fMRI voxel timecourse x(t) and the set of basis functions finds the linear combination of the This is equivalent to an F−test. fMRI timecourses Dept. of Biomedical Engineering, University Hospital SE−581 85 Linköping, Sweden E−mail: olafr@imt.liu.se

y (t)i

Instead a set of basis functions , i = 1...n, can be used.

fMRI timecourse

BOLD model

Analyzing voxels separately does not provide the desired detection sensitivity, and scattered active voxels is a characteristic result.

Canonical Correlation

Presently, the most widely used methods for detecting activity

Ola Friman, Peter Lundberg, Magnus Borga, Jonny Cedefamn and Hans Knutsson, Linköping University, Sweden

MULTIPLE CORRELATION ANALYSIS

y

fMRI timecourse

[1] Friman O et al. (2001) Magn. Reson. Med. 45(2) : 323−330

REFERENCES

w

ρ

CANONICAL CORRELATION ANALYSIS

in BOLD fMRI are the t−test and the ordinary cross−correlation.

y ρ = 0.66 4 (t) y = 0.75 2 ρ y ρ= 0.57 6 (t) y 5 (t) (t) (t) 1 y 3 (t) y 2 1 (t) (t) y 6 5 y y y ρ (t) y (t) = 0.43 4 (t) y 3 (t)

with a Gaussian filter and using the truncated Fourier set yields a correlation of .

SUMMARY

Truncated Fourier set _{Truncated Fourier set}

x w 6 9 5 8 x 2 4 7 1 x x x x x x x t y x 3 x x y1 . . . yn ρ w t y