October 1, 2009 / Vol. 34, No. 19 / OPTICS LETTERS 2933

Three-dimensional fluorescence opticaltomography in small-animal imaging using

simultaneous positron-emission-tomography priors

Changqing Li,* Guobao Wang, Jinyi Qi, and Simon R. CherryDepartment of Biomedical Engineering, University of California, Davis, Davis, California 95616, USA

*Corresponding author: cqli@ucdavis.edu

Received April 14, 2009; revised August 6, 2009; accepted August 24, 2009;posted August 31, 2009 (Doc. ID 109970); published September 22, 2009

We propose three-dimensional fluorescence optical tomography imaging for small animals guided by simul-taneous priors from positron emission tomography (PET). Using a sparsity regularization, the target priorfrom PET images was incorporated into the reconstruction algorithm for optical imaging. The method andalgorithms are validated with numerical simulations and a phantom experiment performed on a combinedoptical-PET scanner based on a conical mirror geometry. © 2009 Optical Society of America

OCIS codes: 170.6960, 170.3880, 170.3890.

While 3D fluorescence optical tomography (FOT) im-aging is limited by relatively low spatial resolution,especially at depth, it is still an attractive, noninva-sive approach for imaging the fluorescence biodistri-bution inside small animals. Multispectral or hyper-spectral measurements [1], accurate forward model-ing [2], and novel reconstruction methods [3] areamong the approaches being explored to improve spa-tial resolution in 3D FOT. It has also been shownthat the spatial resolution of 3D FOT can be im-proved when anatomical images from x-ray computedtomography (CT) or magnetic resonance imaging(MRI) with relatively higher spatial resolution areapplied as a structural prior in the reconstruction ofoptical images [4].

It is common to combine an anatomical imagingmodality, such as CT or MRI, with a functional imag-ing modality, such as PET or optical imaging, withthe widespread adoption of PET/CT being an obviousexample. However, the power of such a combinationin improving the resolution of functional images islimited by position and signal mismatches [5,6]. Thefirst type of mismatch is caused by subject and/or in-ternal organ movements and could be minimized bysystems that permit simultaneous imaging and res-piratory gating. But signal mismatches are unavoid-able because of the dramatic difference between theinformation content of anatomical images and func-tional images [5]. In this Letter, we propose the use ofPET images as prior information for guiding FOT re-construction, in which signal mismatches can be sig-nificantly reduced. For example, if a fluorescent opti-cal probe and PET radionuclide are incorporated intoa single biomolecule [7] or nanoparticle, simulta-neous PET and FOT measurements allow character-ization of the temporal and spatial characteristics ofthese imaging probes with minimal signal mismatch.PET for small-animal imaging is a well-developedtechnology with spatial resolution �1 mm [8], while3D FOT is still under development with spatial reso-lution typically around several millimeters at depth[9]. Incorporating PET information as a prior into op-

tical image reconstruction can potentially improve

0146-9592/09/192933-3/$15.00 ©

the resolution of optical imaging to around 1 mm.This has important applications when the opticalprobe is designed to sense its surrounding environ-ment (e.g., oxygen level, pH, protein–protein interac-tions, and smart optical probes where fluorescence isactivated by the existence of a particular molecule orprotein [10]). In this situation, PET provides funda-mentally different information than optical imaging.PET allows the total probe concentration and distri-bution in the body to be determined, while the fluo-rescent signal adds further information related to theenvironment that the probe finds itself in. The PETdata therefore identifies the locations where an opti-cal signal could arise and can be used as a constraintin the reconstruction. A second application of com-bined PET/optical imaging is to study and validate3D FOT imaging algorithms.

We have developed a conical-mirror-based 3D FOTsystem for small animals [11] that is compatible withsmall-animal PET scanners, allowing simultaneous3D FOT and PET imaging. This system provides datato allow the use of PET priors in 3D FOT reconstruc-tion to be evaluated. The 3D FOT system has beendescribed in detail elsewhere [11]. Briefly, as shownin Fig. 1, the optical system uses a conical mirror, al-lowing simultaneous viewing of the entire surface ofthe animal by an electron-multiplying CCD (EM-CCD) camera. The mirror fits inside the gantry of amicroPET scanner [12] for simultaneous collection ofPET and 3D FOT data. The axial length of the mirroris 6.25 cm with an axial field of view of 6 cm. A650 nm excitation laser collimated to a spot diameterof 1 mm was used. Two motorized mirrors scannedthe laser beam across the object surface. Bandpassfilters, placed inside a filter wheel, were used to se-lect emission wavelengths, and fluorescence signalwas detected by an EMCCD. The microPET scannerand its performance have been described in detailpreviously [12].

The proposed 3D FOT method using PET priors es-timates the 3D optical image x�Rnj�1 defined on thefinite element nodes by the following penalized least-

squares formulation:

2009 Optical Society of America

2934 OPTICS LETTERS / Vol. 34, No. 19 / October 1, 2009

x̂ = arg minx�0

��x�,

��x� = �y − Px�2 + �UwPET

sparsity�x�,

where y�Rni�1 is the optical measurement, ni is thetotal number of optical measurements, and nj is thetotal number of nodes. P�Rni�nj is the forwardmodel of the 3D FOT system. It is computed based onthe forward model of photon propagation in turbidmedia at both excitation and emission wavelengths.In the experiments described below, the diffusionequation is solved with the finite-element method inMATLAB. The surface boundaries of the object beingimaged are obtained by scanning a line pattern laser,and the attenuation and scatter coefficients at the ex-citation and emission wavelengths are assumed to beuniform inside the object. UwPET

sparsity�x� is the sparsityregularization guided by the PET prior, and the hy-perparameter � controls the trade-off between datafidelity and the prior. The PET guided sparsity regu-larization has the following form:

UwPET

sparsity�x� = �j=1

nj

wj�xj�,

where wj is the weighting factor that is determinedfrom the PET image. Here we set wj=1 if the PET ac-tivity at the jth node is below a certain threshold andwj=0 otherwise. Note that this prior simply encour-ages the colocation of the PET tracer and the opticalprobe but does not assume any relationship betweenthe PET signal intensity and optical signal intensity.The sparsity penalty is active only outside the regiondefined by the PET signal. We developed anexpectation-maximization shrinkage algorithm tosolve the penalized least-squares optimization prob-lem. The hyperparameter � was selected empiricallybased on reconstructions of simulated data.

A cubic phantom measuring 32 mm�32 mm�29 mm was composed of 1% intralipid, 2% agar,and 20 µM deoxy-hemoglobin to provide realistic op-tical properties. Four capillary tubes, 1 mm inner di-

Fig. 1. (Color online) Schematic of simultaneous 3D FOTand PET imaging system.

ameter and 12 mm long, were embedded inside the

phantom. Two of the targets were filled with 6.5 µMDiD fluorescence dye solution. All four targets con-tained �18F�-fluoro-2-deoxy-D-glucose (FDG) at an ac-tivity level of 100 µCi, uniformly distributed insidethe capillary tubes. The cubic phantom was placed atthe center of the conical mirror on a transparentstage for simultaneous PET and 3D FOT imaging.The excitation laser scanned the front surface of thephantom [Fig. 2(a)]. Measurements at a wavelengthof 720 nm were used for FOT [Fig. 2(b)]. 1057 detec-tor nodes were selected on the four side surfaces ofthe phantom. A finite-element mesh with 8690 nodesand 47,581 tetrahedral elements was used for opticalforward model calculation. PET images are shown inFigs. 2(c)–2(e), where the color scale indicates therelative FDG concentration. From the PET images,we obtained capillary tube positions (the nodes in thefinite-element mesh) by segmentation with a thresh-old of 20% of maximum FDG concentration in PETimages (see Figs. 4(a) and 4(b) for the segmented tar-get zone). This radiotracer prior was used in the 3DFOT reconstruction.

The reconstructed 3D FOT images of the physicalphantom are shown in Fig. 3. We studied two differ-ent priors. In the first experiment, we used the PETimage of the tubes containing the two fluorescencetargets (T1 and T2) to construct the prior, i.e., no mis-match. The resulting reconstructed 3D FOT imagesare shown in Figs. 3(c) and 3(d). In the second experi-ment, we included targets T3 and T4, where therewas only FDG contrast and no fluorescence contrast,i.e., mismatch between PET and FOT images. The re-sulting reconstructed 3D FOT images are shown inFigs. 3(e) and 3(f). We did not observe any degrada-tion or artifacts introduced from the mismatchedprior. From Fig. 3, we see that, without the prior, thetwo fluorescence targets are detected in the right lo-cation but with poor spatial resolution FWHM5.2 mm). When the prior was added, the detected tar-gets have better spatial resolution, as shown in Figs.3(c)–3(f). In both cases the FWHM is improved to

Fig. 2. (Color online) (a) Cubic phantom with four embed-ded capillary targets (T1–T4); (b) fluorescence measure-ments from different surfaces of the cubic phantom placedin the conical mirror. PET images showing (c) transverse,

(d) sagittal, and (e) coronal sections.

October 1, 2009 / Vol. 34, No. 19 / OPTICS LETTERS 2935

2.8 mm. Some artifacts are observed at the boundaryof the reconstructed image, and these are likely dueto measurement noise.

For numerical simulation, a cubic phantom withthe same dimensions and optical properties as thereal phantom was used, with two embedded capillarytargets at the same location as T1 and T2 in Fig. 2.The target shape was derived by thresholding thePET image from the measured phantom data. Thefluorescence concentrations in the two targets were100 and 75 [arbitrary units, Figs. 4(a) and 4(b)]. Thebackground has no fluorescence. The simulation mea-surements were obtained by forward modeling, and20% Gaussian noise was added to each measure-ment. Just as in the phantom data, 20 illuminationpositions were simulated on the front surface, and1057 measurement nodes were simulated on the fourside surfaces. The same finite-element mesh wasused. The reconstructed 3D FOT images with and

Fig. 3. (Color online) Reconstructed 3D FOT images fromthe phantom experiment showing sagittal sections (toprow) and transverse sections (bottom row); (a), (b) withoutprior, (c), (d) with correct prior, and (e), (f), with redundantprior. The color bar indicates relative fluorescenceconcentration.

Fig. 4. (Color online) Numerical simulations: True fluores-cence concentration for cubic phantom in (a) sagittal and(b) transverse sections. Reconstructed images (c), (d) with-out prior and (e), (f) with prior. The color bar indicates rela-

tive fluorescence concentration.

without the prior at iteration 2000 are plotted in Fig.4. Without the sparsity regularization, the mean re-constructed fluorescence concentrations in the twotargets were 32.8 and 28.5. With the sparsity regu-larization, these values are 78.1 and 53.4, closer tothe true values. The improved resolution of the im-ages that incorporate the PET prior leads to im-proved quantification by reducing partial volume ef-fects.

In summary, we have proposed a method for 3DFOT reconstruction that utilizes prior informationfrom PET. A reconstruction algorithm has been devel-oped and evaluated with numerical simulation and aphantom experiment. The results demonstrate that aPET prior can improve the spatial resolution andquantitative accuracy of 3D FOT images. The advan-tage of a PET prior, compared with an anatomicalprior from CT or MRI, is the potential to minimizethe signal mismatch between the two modalities andhence improve the effect of prior information. Thiswill be of particular value for quantifying fluorescentprobes whose signal depends on the biochemical en-vironment or on an interaction with specific proteins.The quality of the reconstructed 3D FOT imagescould be further improved by using more illumina-tion positions, optimizing illumination patterns, andemploying measurements at multiple wavelengths.Future studies will also include the development ofdual PET-FOT probes and their applications for invivo mouse imaging.

References

1. A. J. Chaudhari, F. Darvas, J. R. Bading, R. A. Moats,P. S. Conti, D. J. Smith, S. R. Cherry, and R. M. Leahy,Phys. Med. Biol. 50, 5421 (2005).

2. Z. Yuan, X. H. Hu, and H. B. Jiang, Phys. Med. Biol.54, 65 (2009).

3. S. Ahn, A. J. Chaudhari, F. Darvas, C. A. Bouman, andR. M. Leahy, Phys. Med. Biol. 53, 3921 (2008).

4. A. H. Barnett, J. P. Culver, A. G. Sorensen, A. Dale,and D. A. Boas, Appl. Opt. 42, 3095 (2003).

5. C. Comtat, P. E. Kinahan, J. A. Fessler, T. Beyer, D. W.Townsend, M. Defrise, and C. Michel, Phys. Med. Biol.47, 1 (2002).

6. J. Culver, W. Akers, and S. Achilefu, J. Nucl. Med. 49,169 (2008).

7. W. B. Edwards, B. Xu, W. Akers, P. P. Cheney, K.Liang, B. E. Rogers, C. J. Anderson, and S. Achilefu,Bioconjugate Chem. 19, 192 (2008).

8. Y. F. Yang, Y. B. Wu, J. Y. Qi, S. S. James, H. N. Du, P.A. Dokhale, K. S. Shah, R. Farrell, and S. R. Cherry, J.Nucl. Med. 49, 1132 (2008).

9. G. Zacharakis, J. Ripoll, R. Weissleder, and V.Ntziachristos, IEEE Trans. Med. Imaging 24, 878(2005).

10. R. Weissleder, C. H. Tung, U. Mahmood, and A.Bogdanov, Jr., Nat. Biotechnol. 17, 375 (1999).

11. C. Q. Li, G. S. Mitchell, J. Dutta, S. Ahn, R. M. Leahy,and S. R. Cherry, Opt. Express 17, 7571 (2009).

12. Y. Yang, Y. C. Tai, S. Siegel, D. F. Newport, B. Bai, Q.Li, R. M. Leahy, and S. R. Cherry, Phys. Med. Biol. 49,

2527 (2004).