DCM for ASL data Introduction CONCLUSION ASL ...

4 downloads 0 Views 4MB Size Report
We propose a new version of the standard fMRI-DCM generative model further extended by the .... BOLD comes mainly from veins and their surrounding tissue.
Dynamic Causal Modeling for Arterial Spin Labeling data 1

1

2

1

1

1

Martin Havlicek , Alard Roebroeck , Karl Friston , Dimo Ivanov , Anna Gardumi and Kamil Uludag 1 Dept. of Cognitive Neuroscience, Faculty of Psychology, Maastricht University, Netherlands, 2 The Wellcome Trust Center for Neuroimaging, UCL, WC1N 3BG, UK,

[email protected]

Introduction

ASL basics TR

TR

Control

Label

Mz(blood)

Control

Labeled

Label

Induc. Signal

ASL Model

Full model inversion

Modulation

BOLD

Input

ASL signal

R3

0.3

-0.5

R3

Input 0.2

R1

0.4

R2

-0.2

Fig. 5. 50 Monte Carlo (MC) simulations of 3 node network with a similar connectivity structure as shown on the left. were generated. Always having one driving input; one backward connection with an opposite sign; and one modulatroy connection. The final ASL and BOLD time courses were modeled assuming regionspecific variability of hemodynamic response; sampled with TR = 2 sec; total length 610 sec; and SNR = 1.

Modulation

-0.5

0.2

dHb

Blood Flow

Fig. 1. ASL uses spatially selecASL time course tive magnetic inversion of inflowing arterial blood as a method to label blood flow. The MR signal from inverted blood is made negative relative to uninverted blood. When the labeled blood reaches the tissue, it attenuates the signal from the Time (TR) image of that tissue. Subtraction of labeled images from a control images gives a measure of the amount of label which flowed into the tissue. This is proportional to the CBF signal. The measured CBF signal comes mainly from arterioles, capilaries and brain tissue, whereas BOLD comes mainly from veins and their surrounding tissue.

Simulated data

R1 Neural Signal

t

Inversion Time

Simulated

Blood Volume

∆M CBF

Labeled

Label

Full Arterial Spin Labeling Model based DCM Hemodynamic Model

CBF

Control

DCM for ASL data Connectivity Model

Labeled

-

Wait

Dynamic causal modeling (DCM) [2] is a standard method to assess an effective connectivity from functional brain data. While blood oxygenation level dependent (BOLD) contrast fMRI data are the most commonly used in DCM analysis, there are other fMRI measures such as arterial spin labeling (ASL) that could be used as well: ASL has the potential to provide a better localization of the functional signal to the site of neural activity as compared to BOLD contrast [1]. Both cerebral blood flow (CBF) and BOLD signals can be derived from ASL data. Our primary goal is to establish a DCM for the evaluation of effective connectivity from ASL data. We propose a new version of the standard fMRI-DCM generative model further extended by the genrative model of ASL data. Our secondary goal is to use both CBF and BOLD components of ASL data to optimize regionspecific hemodynamic parameters and thereby prevent confounding hemodynamic variability (within and across brain regions) with estimated neuronal dynamics (i.e. connectivity parameters).

R2

-0.5

Additive Model Baseline × Blood Flow

Driving Signal

Fig. 2. In order to enable DCM for ASL data, the standard BOLD-DCM [2], which consists of neuronal and hemodynamic model, is further extended with a model of ASL signal. This model is represented by three additive components: the modeled BOLD signal (purple line); the modulated CBF signal (green line); and the ASL baseline (black line). The CBF and the BOLD signals are provided by the hemodynamic model. Note that besides this clear extension, we also employ an extended hemodynamic model [4], which is briefly described below. The estimation of ASL-DCM is performed using a standard Variational Bayesian EM algorithm. Additionally, having both CBF and BOLD components we are able to estimate important region-specific hemodyamic parameters that are then used as a priors (in Bayes optimal fashion) in ASL-DCM. In this case, we consider a flexible neuronal model using basis functions.

Flexible neuronal model for ROI-specific estimation of HDM parameters

Basis Functions

Magnitude (a.u.) 0 10

20

30

40

Time (sec)

0.4

A

42/50

0.04

0

10

20

30

A

Neuronal response B

CBF response BOLD response

“Purely hemodynamic...” 0 0

10

20

30

40

Time (sec)

Rest

CBF response BOLD responses

CBF response BOLD responses

40

C

Rest

Rest

D Rest

Rest

...

BOLD

26 s

14 s

B

C

D

CBF

conjunction BOLD & CBF 8

20

8

15

6

6

10

4

4

5

2

2

0

0

0

Fig. 7. The task consisted of four conditions organized in blocks, interleaved with resting periods, which were randomly repeated 5 times per run. Each block included 11 pseudorandomized events represented by visual stimuli in left and right visual field. The subject was fixating centrally and pressed the button using the index finger of his left or right hand according to the side of visual stimuli (conditions A and B). If the fixation cross turned red, the subject responed with the hand contralateral to the side of stimuli (condition C and D).

M1 Identified network using ASL-DCM:

Fig. 8. The functional ASL data were preprocessed (realigned; coregistered; slice-timing corrected; smoothed with 5 mm FWHM Gaussin kernel) and analyzed using a modified SPM8 functions (considering a full ASL GLM model that besides the main ASL regressors [3] included also modeling of low frequency confounds and AR(1) noise component). Statistical parametric maps were obtained for both BOLD and CBF components (p