fMRI Noise Problems and Methods to Mitigate

1
fMRI Noise Problems and Methods to Mitigate

• Presented by
• Clay McCreary

2
Introduction

• Noise can affect an fMRI imaging session in two
  ways
• Creates artifacts in the image
• Strong acoustic noise interferes with the audio
  stimuli (talking to the patient) used in imaging
  experiments, and creates a harmful, annoying
  environment for both the patient and the
  staff4.

3
Creating Artifacts

• Drift Noise
• Reference 1 defines drift as low frequency
  signal that varies slowly across the whole period
  of the acquisition of the data caused by12
• Gross head motion
• Cerebrospinal fluid (CSF) pulsations
• Physiological fluctuations
• Local Magnetic susceptibility changes
• Image Noise
• Reference 1 has determined through observation
  that the noise in fMRI has the characteristics of
  fractional noise and exhibits long range
  autocorrelation in time.

4
Fractional Noise

• Sometimes referred to as Pink Noise3
• Signal or process with a frequency spectrum such
  that the power spectral density is proportional
  to the reciprocal of the frequency3.
• There is equal energy in all octaves (or similar
  log bundles)3.
• White noise has equal energy per hertz3
• In terms of power at a constant bandwidth, 1/f
  noise falls off at 3dB per octave3.

5
Modeling Drift Noise

• Drift noise is assumed to vary slowly with large
  scales1.
• A discrete wavelet transform is used to model
  drift noise1
• Fine scale wavelet coefficients will be zero
  since drift noise is low frequency and thus does
  not vary greatly over a short period of time.

6
Modeling Image and Acoustic Noise

• The image noise in the fMRI time series obtained
  under the resting or null conditions exhibit
  long-range autocorrelation in time and 1/f-like
  spectral properties (fractional noise)1.
• Acoustic noise is AWGN.
• Generated by the rapid switching of the scanner
  coil to produce the strong magnetic field gradient

7
Methods to Solve Noise Problem

• Modified General Linear Model and Bayesian
  Estimator1
• Blind Dereverberation for fMRI Noise Based on
  SIMO Linear Prediction Method4
• LMS-based Active Noise Cancellation Methods for
  fMRI Using Sub-band Filtering5

8
Modified General Linear Model and Bayesian
Estimator1

• Discrete wavelet decomposition is applied to the
  signal.
• yi ßbi fi ni, i 1, . . . , N (assumed LTI)
• yi is the fMRI signal at a specific voxel
• ß is a scalar estimation of the contribution of
  the desired signal to yi
• fi is the contribution of drift to yi
• ni is the contribution of noise to yi

9
Modified General Linear Model and Bayesian
Estimator1 (cont)

• Since fi is present only in large scale wavelets,
  finer scales are used for this calculation
  removing this term.
• Applying an orthonormal wavelet process
  decorrelates the noise and the wavelet
  coefficients are normally distributed with zero
  mean and identical variances at the same
  levels1.
• Bayesian estimator accurately determines the
  noise covariance matrix thus ß is accurately
  determined.

10
Blind Dereverberation for fMRI Noise Based on
SIMO Linear Prediction Method4

• Active noise cancellation (ANC) is used to
  mitigate acoustic noise
• Room reverberation degrades the performance of
  the ANC
• This method models the room as a single input
  multiple output (SIMO) system using two
  microphones to capture some of the reverberation
  from the room.

11
Blind Dereverberation for fMRI Noise Based on
SIMO Linear Prediction Method4(cont)

• An algorithm based on a prediction method to
  reduce the reverberation from the two
  microphones output is used on the SIMO
  system4.
• The output of the microphones correlates with the
  entire room reverberation.
• The room reverberation is predicted and cancelled
  without the room transfer function.

12
LMS-based Active Noise Cancellation Methods for
fMRI Using Sub-band Filtering5

• Using a least mean squared (LMS) finite impulse
  response (FIR) filter over the entire band of
  acoustic noise produced by the fMRI will result
  in a heavy computational load and significant
  time delay.

13
LMS-based Active Noise Cancellation Methods for
fMRI Using Sub-band Filtering5(cont)

• Dividing the acoustic band into subbands and
  implementing LMS-FIR filters on each subband then
  combining them using a stacking method allows
  for less computational load and nearly real time
  results.
• Since the subband method is essentially
  delayless it provides a more effective means of
  ANC.

14
References

• 1 Huaien Luo and Sadasivan Puthusserypady,
  Analysis of fMRI Data With Drift Modified
  General Linear Model and Bayesian Estimator,
  IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL.
  55, NO. 5, pp. 1504-1511, May 2008
• 2 T. Kim, L. Al-Dayeh, and M. Singh, FMRI
  Artifacts Reduction Using Bayesian Image
  Processing, IEEE TRANSACTIONS ON NUCLEAR
  SCIENCE, VOL. 46, NO. 6, pp. 2134-2140, DECEMBER
  1999

15
References (cont)

• 3 Wikipedia, Pink Noise, Retrieved from the
  world wide web, June 7, 2008 from
  http//en.wikipedia.org/wiki/1/f_noise
• 4 Hua Bao, Issa M.S. Panahi, Richard Briggs,
  Blind Dereverberation for fMRI Noise Based on
  SIMO Linear Prediction Method, Proceedings of
  the 29th Annual International Conference of the
  IEEE EMBS Cité Internationale, Lyon, France, pp.
  2831-2834, 2007

16
References (cont)

• 5 Ali A. Milani, Student Member. Issa Panahi,
  Richard Briggs, LMS-based Active Noise
  Cancellation Methods for fMRI Using Sub-band
  Filtering, Proceedings of the 28th IEEE EMBS
  Annual International Conference New York City,
  USA, pp. 513-516, Aug 30-Sept 3, 2006