Abstract Arterial Spin Labeling (ASL) is a noninvasive method for

quantifying Cerebral Blood Flow (CBF). The most common approach is to alternate between tagged

and non tagged MRI images. Averaging is then performed, in order to detect weak

magnetization differences among control and labeled images. A new method is proposed, in which the magnetization

estimation problem is formulated in a Bayesian framework. Spatial-temporal priors are used to deal with the ill-posed

nature of the reconstruction task. The rigid alternating tagging strategy constraint imposed in the

traditional ASL methods is no longer needed. Tested with synthetic and real data, the algorithm proposed has

shown to outperform the traditional averaging methods used.

Bayesian perfusion estimation with PASL-MRI

RecPad2010 - 16th edition of the Portuguese Conference on Pattern Recognition, UTAD University, Vila Real city, October 29th

M. L. Rodrigues, P. Figueiredo and J.Miguel SanchesInstitute for Systems and Robotics / Instituto Superior Técnico

Lisboa, Portugal

Experimental Results-To test the algorithm, a mask was created, similar to the human brain, with two different regions (white and gray matter).

-A sequence of Monte-Carlo tests was performed, with the following values:σ=1, F=1000, ΔM(gray matter)=0.5 and ΔM (white matter)=1;-The values obtained pre-processing were: SNR(F)=80.0228dB and SNR(ΔM)= -2.20135dB.

-The results reveal a major improvement in both the final SNR of the image (≈3dB) and the mean error ( 8%).-Applied to real data, images revealed less influence of noise and smoothing of areas of the same intensity. Also, better definition on the contours.-These are important results, that allow the decrease of long acquisition times necessary to acquire at multiple TI, without compromising image quality.

IntroductionArterial spin labeling:

1.Arterial blood passing through the carotid is labeled with an inversion pulse;

2. After an Inversion Time (TI), the image is acquired;3. The procedure is repeated, only this time no inversion

pulse is applied.4. Control image is acquired;5. Subtracting the control and labeled images, a difference of

magnetization is obtained, which is an indicator of CBF.

)|()|(log)|(log),,,|(logmaxarg)ˆ,ˆ,ˆ(

vDpvFpvMpvDFMYpDFM

Results – Comparison of the 3 methods

Pair Wise subtraction

Surround Subtraction

Algorithm

SNR(ΔM)(dB) 12.221 12.2796 15.6200

ISNR(ΔM)(dB) 14.4231 14.4810 17.8214

Mean Error (%) 23.40 23.07 15.12

Problem Formulation-The algorithm proposed is designed in a Bayesian framework, with the following observation model:

-Y :3 D matrix (n x m x l) (a stack of l images of n x m pixels);- F (n x m) is the base morphological MRI image;- D (n x m x l) represents the Drift;- ΔM (n x m): magnetization difference measured; -Γ~N(0,σy

2)(n x m x l)Additive White Gaussian Noise (AWGN);- v (l x 1) contains the labeling marks indicating which image

among the sequence is labeled.-The estimation of ΔM given the observations Y and the vector v is a ill-posed problem and prior information is needed to regularize the solution.-The Maximum A Posteriori (MAP) estimation problem can be formulated as:

MvDFYFigure 3: Processed images of synthetic data using the three methods

Figure 4: Processed images of real data, using the three different methods: Top left - Pair-wise subtraction; Top right - surround subtraction; Bottom three - algorithm using different parameters

Figure 1: Schematic of the Arterial Spin Labeling procedure

Figure 2: Sampling strategy of PASL