Functional Brain Imaging combining DOI and fMRI


1.  About Diffuse Optical Imaging

Diffuse Optical Imaging (DOI) is a relatively new method used to image blood volume and oxygen saturation in vivo.  It uses near infrared light and has the advantages of being low cost and portable.  The properties of near infrared light in biological tissue make diffuse optical imaging techniques very successful.  The absorption coefficient (µa) depends on the total hemoglobin concentration and oxygenation level within the tissue; therefore, calculating µa provides useful information about the physiological conditions of the tissue.  For instance, during the last few years DOI has been tested for application to breast cancer imaging, brain oxygenation, brain trauma, and brain function.


2.  Motivation

Various methods are known for producing functional brain images.  Functional Magnetic Resonance Imaging (fMRI) provides good spatial resolution but poor temporal resolution whereas DOI exhibits excellent temporal resolution but is penalized by the highly scattering superficial head layers that make it difficult for light to reach the brain.  Therefore, the DOI signal scattered back from the brain and detected by the optodes placed on the scalp is fairly weak.


Intuitively, a multi-modality method should provide more accurate functional images of the brain because it combines the strengths of diverse methods.  The current project aims to thoroughly explore the potential of a dual-modality system that uses MRI to spatially guide DOI.


3.  Dual-modality method and preliminary results

The idea is to use fMRI to localize the region of the cortex where there is activity during the performance of a task (e.g. finger tapping, visual stimuli, etc.) and to linearly combine the data acquired using fMRI with that gathered simultaneously by DOI.


Using a soft spatial prior provided by fMRI we add information about the position of the activation region in the brain.  By weighting such prior we can make it more or less relevant in the reconstruction process.  For example, if we believe that DOI can localize the activation region by itself, we can assign a zero-weight to the fMRI prior.  On the other hand, if the activation region is in deeper layers and therefore hardly detectable by our optical system, we can increase the contribution of the spatial prior in the reconstruction process.


Unfortunately, the soft prior added to the Tikhonov functional is not strong enough to significantly improve the restored contrast image. Since the problem is extremely underdetermined (84,864 unknowns and 114 measurements), the restoration algorithm is more likely to find activation in the regions where the detectors are most sensitive to absorption changes. Therefore, activation is typically found at the superficial layers (note that light intensity decays exponentially with depth).


Figure 1 shows the head model and the locations of maximal sensitivity to absorption in the cortex (white crosses) of each optode (numbered in black) for a given coronal slice.


Figure 1


In order to improve localization of the activated region it is necessary to reduce the number of voxels involved in the restoration process. A simple way to acquire information on the inactive region (such as scalp and skull tissue types, where brain activation will not induce absorption coefficient changes) is to perform an MR anatomical scan and assign zero value to the voxels corresponding to scalp and skull in the imaging matrix calculated from DOI. We call the introduction of such a hard constraint to the DOI forward model the scalp-skull prior (or Hard Brain in figure 2a and figure 2b).


Figure 2 and figure 3 demonstrate the improvement of the dual-modality method over each technique alone: figure 2a and figure 2b show coronal slices of the brain reconstructed using the Tikhonov inverse method and a hard prior (the second column shows the restoration using the scalp-skull prior, whereas the third column shows the result calculated by adding the cortical prior as in [1]).  Figure 2a presents an example of two coronal slices where the use of the scalp-skull prior produces better images than using the cortical prior; figure 2b, on the other hand, shows a case where the use of the cortical prior generates more accurate reconstructions.  The columns and rows in figure 2 correspond to different restoration modality and coronal slices, respectively. The first column shows the true simulated activation region, the second column shows the restoration obtained using the scalp-skull prior, and the third column shows the results calculated using the cortical prior.



Figure 2a


Figure 2b


Figure 3 compares a simple Tikhonov reconstruction with a Tikhonov restoration using an fMRI hard prior (i.e. forcing activation to be found in the location identified by the fMRI data). The center column of the figure shows how powerful such a prior is and how much information is therefore lost.  Ideally we would like to find a compromise that uses as much information as possible and preserves it through the reconstruction process but at the same time finds the activation region position with the greatest possible accuracy.


Figure 3


4.  Summary

The above data showed that using a spatial prior clearly improves the localization of the activation region in the brain.  The tests were performed using only one temporal frame.  Combining all the temporal frames will express more accurately the temporal evolution of the activation due to task performance and the localization of the activation region in the brain.  The localization, in particular, was improved by the contribution of the MRI data.


The use of a scalp-skull prior (i.e. adding information on locations of non active regions) instead of a hard cortical prior (i.e. forcing sensitivity to µa changes to be only in the cortex) has two main advantages: it decreases the prior information used in the restoration process and it reduces the computation time required to calculate the prior.  This is because computing the hard prior necessitates acquiring a full anatomical MR image of the subject head and segmenting it into the various tissue types, whereas computing the scalp-skull prior involves the performance of a preliminary MR anatomical scan without segmentation that is extremely less computationally expensive.


5.  References

[1] D. Boas and A. Dale, “A simulation study of MRI guided cortically constrained diffuse optical tomography of human brain function” (2004)




Anna Custo          custo[at]

David Boas           dboas[at]

Eric Grimson          welg[at]

William Wells       sw[at]

John Fisher           fisher[at]




Back to the Medical Vision Group page.

Back to the MIT CSAIL page.

Last updated July 29, 2004.