The FM were composed of four sources of germanium-68 (74 kBq; diameter, 1 mm; and length, 0.5 mm; Isotop Product Laboratories, Valencia, CA, USA) originally designed for PET-CT co-registration, included in the center of four Plexiglas cubes of 1 × 1 × 1 cm. A spot of 2 mm diameter was drawn with standard white liquid corrector (Tipp-Ex®, Bic, Clichy, France) on top of each Plexiglas cube, exactly above the ⁶⁸Ge sources. The cubes were then glued permanently to a custom made transparent Plexiglas mouse supporting plate at four corners close to the position of the mouse on the plate (Figure 1A).

Animal experiments were performed under an animal use and care protocol approved by the animal ethics committee and conducted in accordance with Directive 2010/63/EU of the European Parliament. Six female nude mice (body weight of approximately 25 g) were obtained from Elevage Janvier (Le Genest Saint Isle, France) and received a subcutaneous injection in the flank of 10⁶ PC12-multiple endocrine neoplasis syndrome type 2A (MEN2A) cells 20 . The mice were anesthetized by continuous gaseous anesthesia (1–2% isoflurane in O₂) and imaged sequentially by fDOT and PET. The near-infrared (NIR) fluorescent dye Sentidye® (20 nmol; Fluoptics, Grenoble, France) was injected 3 h before starting the fDOT acquisition at a volume of 100 μL. FDG (7,400 kBq in 100 μL; Flucis, IBA, France) was administered 1 h before the PET scan. Each mouse underwent a 20-min fDOT acquisition followed by a 30-min PET acquisition. The anesthetized mice were transferred from the fDOT to PET scanner by means of the mouse supporting plate, while great care was taken to avoid movement of the animal in regard to its support. The contact of the ventral side of a nude mouse with the Plexiglas surface of the mouse holder is sticky enough to ensure that the mouse is not moving when transferred between scanners located at the same room.

Images were acquired in the 3-D optical imager TomoFluo3-D (Cyberstar, Grenoble, France) 7 . To obtain the position of the FM, two optical images covering both the subject and the FM were acquired with the mouse placed in prone position: (1) a planar white light image recorded from a camera snapshot yielding the x and y coordinates (Figure 1B), and (2) an image of the 3-D surface of the animal, acquired by rapid consecutive camera shootings during axial scanning with an inclined green planar laser of the TomoFluo3-D, yielding the z coordinates (Figure 1C). One image was recorded for each laser position during laser scanning of the animal in the axial direction, and all images were then combined into a single image 7 . A surface image of the animal was then reconstructed from the intersection curve between the surface of the animal and the surface of the supporting plate by triangulation 7 .

The fDOT image was obtained by a 20-min scan of a defined volume of interest covering the tumor using excitation by a 680-nm laser on the anterior side of the animal 7 and recording with a CCD camera fixed above its dorsal side. The scanning grid consisted of 7 × 6 sources in steps of 2 mm, and the detection area was 15 × 13 mm². A 2 × 2 binning was applied, and the mesh volume corresponding to the detection area was mathematically discretized in voxels of 0.67 × 0.67 × 1 mm³ size to build the reconstruction mesh volume. Finally, the inverse problem of the tomographic reconstruction was solved with the algebraic reconstruction technique 7 .

A 30-min scan was acquired in a Focus 220 MicroPET scanner (Siemens, Knoxville, TN, USA). Image acquisition and reconstruction used the MicroPET Manager Software (Siemens-Concorde Microsystems) based on a filtered back-projection algorithm. The attenuation correction was based on the segmentation of the emission map 21 ; the dimensions of reconstruction volumes were 256 × 256 × 95 with a voxel size of 0.475 × 0.475 × 0.796 mm³. The counts were decay-corrected and expressed in Bq/cm³.

As the three optical images (white light image, surface image and fDOT reconstruction) were acquired in the same spatial referential, the transformation matrix between the fDOT image and the white light image T _fDOT−photo is calculated using the intrinsic parameters of fDOT. Hence, the T _fDOT-PET transformation matrix for fDOT to PET co-registration is a product of T _fDOT-photo and T _photo-PET :

T f D O T − P E T = T f D O T − p h o t o × T p h o t o − P E T

The co-registration (T _photo-PET ) method was processed in four steps:

1. Detection of the optical planar (x, y) FM coordinates.

The Tipp-Ex® spot drawn on the top of each Plexiglas cube helped visualize the planar position of the FM in optical images. Four square-shaped regions of detection (ROD) were assigned onto predetermined positions in the planar white light image (Figure 1B). Each ROD had a size of 6 × 6 mm, corresponding to 30 × 30 pixels in the concatenated mouse photograph, i.e., three times larger than the Tipp-Ex® spots' dimensions in order to obey the Nyquist-Shannon sampling theorem 22 while avoiding parasite signals from the mouse body. The first step consisted in the automatic detection of the x and y coordinates of the FM based on the maximal intensity inside the corresponding RODs (Figure 1B). Three image preprocessing steps were then performed successively: (1) filtering with a 3 × 3 median filter that eliminated most of the noise present in the RODs, (2) high-pass thresholding of image pixel intensities at a threshold value of 90 % of the maximum intensity and (3) application of a Deriche's recursive Gaussian filter 23 in order to center the gradient intensity change in the images of the Tipp-Ex® spots. Following these three steps, the coordinates of the local maximum intensity in each ROD coincided with the center of the FM signal given by the Tipp-Ex® spots in the planar white light image and assigned positions (x _opt, y _opt) of the FM.

2. Detection of the optical altitude (z) FM coordinates.

Since the optical surface image (Figure 1C) and the planar white light image are concatenated in the same orientation and the same pixel size, the (x, y) coordinates in both images correspond directly. The intensity values of the optical surface image representing the distance between the upper surface of the FM and the supporting plate were measured at position x _opt, y _opt to yield the z0_opt value of the upper surface of the FM. Altogether, the combination of the optical surface image and the planar white light image allowed assigning full 3-D coordinates (x _opt, y _opt, z0_opt) to each FM in the optical image.

3. Detection of PET FM coordinates.

Four 3-D RODs of 9 mm³ (dimensions three times larger than the dimension of FM signal in PET image) were defined in the PET volume image (Figure 1E). Following completion of the same image preprocessing steps as for the detection of the optical planar coordinates, the local maximum was detected in each ROD to yield the coordinates (x _PET, y _PET, z _PET) of the FM in the PET volume image. Since the optical and PET signals from the FM do not coincide in the z dimension (i.e., the optical signal is on top of the Plexiglas cube, and the PET signal inserted inside the cube), a distance dz was added to account for translation in the z direction after the calculation of the rigid transformation matrix between the optical and PET image.

4. Transformation from the mouse photograph to the PET volume.

A rigid transformation with translation and rotation was applied to co-register the optical and PET coordinates of the FM. With Po = {Po ₁, Po ₂, Po ₃, Po ₄} and Pp = {Pp ₁, Pp ₂, Pp ₃, Pp ₄} being the four FM positions in the optical and the PET volume images, respectively, the translation T and rotation R were defined as Equation 2:

P p = R ∗ P o + T

The algorithm to compute the transformation R, T used the singular value decomposition (SVD) 24 approach to find the least square error criterion (Equation 3),

∑ i = 1 N E = ∑ i = 1 N P p i - R P o i + T 2 , where N = 4 is the number of FM .

The point sets {Pp _i } and {Po _i } were imposed the same centroid for calculating rotation:

P p ¯ = 1 N ∑ i = 1 N P p i P ^ p i = P p i - P p ¯ P o ¯ = 1 N ∑ i = 1 N P o i P ^ o i = P o i - P o ¯

Rewriting and reducing Equation 3:

∑ i = 1 n E = ∑ i = 1 N P ^ p i - R ^ P ^ o i 2 = ∑ i = 1 N P ^ p i T P ^ p i + P ^ o i T P ^ o i - 2 P ^ p i T R ^ P ^ o i

Transformation was expressed as Equation 6:

T ^ = P p ¯ - R ^ P o ¯

The R ^ , T ^ is the optimal transformation that maps the set {Pp _i } to the {Po _i }. Equation 5, also known as the orthogonal Procrustes problem 24 25 , is minimal when the last term is maximal.

The optical images, being proportionally larger than the PET image due (1) to the difference in the pixel size between the optical planar image (0.21 × 0.21 mm) and the PET image (0.47 × 0.47 mm), and (2) to the parallax induced by the camera detecting the Tipp-Ex® spots from the upper surface of the cube, a scaling factor was applied by calculating the average distance of the points in the x and y axes between the two modalities as (do _x , do _y ) and (dp _x , dp _y ):

d o x = P o 2 ( x ) - P o 1 ( x ) + P o 4 ( x ) - P o 3 ( x ) / 2 d p x = P p 2 ( x ) - P p 1 ( x ) + P p 4 ( x ) - P p 3 ( x ) / 2 d o y = P o 3 ( y ) - P o 1 ( y ) + P o 4 ( y ) - P o 2 ( y ) / 2 d p x = P p 3 ( x ) - P p 1 ( x ) + P p 4 ( x ) - P p 2 ( x ) / 2

A 2 × 2 scaling matrix was then built:

S = S x 0 0 0 S y 0 0 0 s z

where S x = d o x d p x , S y = d o y d p y and S z = 1 (i.e., no scaling in the z direction).

After correcting for distance dz, the final transformation matrix T _photo-PET for the optical photograph to the PET volume was:

T p h o t o – P E T = R ^ S 11 R ^ S 12 R ^ S 13 T ^ x R ^ S 21 R ^ S 22 R ^ S 23 T ^ y R ^ S 31 R ^ S 32 R ^ S 33 T ^ z

where RS are the rotation matrix elements R multiplied by the scaling matrix elements S of Equation 8.

Applying the matrix in Equation 9 to the optical photography (Figure 1F) aligned the 3-D PET image (Figure 1G) to yield the co-registered bimodal image shown in Figure 1H.

The general outline of the co-registration method is depicted schematically in Figure 1. Five steps are implemented successively: (1) registration of (x _opt, y _opt, z0_opt), i.e., the x, y and z coordinates of the optical FM using both the white light (Figure 1 B) and optical surface (Figure 1C) images; (2) calculation of T _fDOT-photo , the transformation matrix from the fDOT reconstruction image to the whole body image; (3) registration of (x _PET, y _PET, z _PET), i.e., the x, y and z coordinates of the FM in the PET images (Figure 1D,E); (4) calculation of T _photo-PET , the transformation matrix from optical to PET volumes (Figure 1H); and (5) calculation of T _fDOT-PET , the transformation matrix for the co-registration of the fDOT volumes with the PET volumes.

The complete image processing method was written in C++ and integrated in Python with a user-friendly interface within the Brainvisa image processing software (http://brainvisa.info/index_f.html) 26 . Input files include (1) the mouse photograph, (2) the optical surface scan, (3) a header file containing the position of the fDOT image relative to the camera's field of view, and (4) the PET volume image. With this user-friendly toolbox, three transformation matrices (T _fDOT-photo , T _photo-PET , and T _fDOT-PET ) are calculated and sent as outputs. The program is completed in 20–30 s on a Dell Precision^TM T7500 (Dell SA, Montpellier, France) workstation with a 64-bit Intel® Xeon® quad-core processor (Intel Corporation, Montpellier SAS, France).

The fDOT-to-PET co-registration method was applied to six female nude mice bearing xenografts tumor of PC12-MEN2A cancer cells that mimic a human medullar thyroid carcinoma 20 . The diameter of tumors ranged from 4.5–8 mm, corresponding to tumor weights of 50–270 mg. Each mouse received two injections; the first injection was the fluorescent probe Sentidye®, a probe that passively accumulates in tumors by virtue of the enhanced permeability and retention effect, and the second one is FDG. Each mouse was imaged sequentially with fDOT and PET.

There was a strong correlation between the volume of the tumor and the total uptake of FDG, but no correlation between the tumor volume and the concentration of FDG in the tumor expressed in percent of injected dose per cubic centimeter. In other words, the radioactivity concentration remained independent of the tumor volume. The fDOT signal also increased with tumor size but plateaued for tumors larger than 150 mg, corresponding to a tumor with a diameter larger than 6 mm. Again, no correlation was found between tumor volume and the concentration of the fDOT signal in the tumor area.

The fDOT-PET fused images (Figure 2) showed the localization of the optical probe with respect to the glucose consumption of the tumor. After the RODs had been defined for the mouse holder as indicated above, fDOT and PET images were co-registered automatically and the localization of FDG and Sentidye® uptakes were compared directly. Tumor volumes measured based on PET-FDG uptake ranged from 53–271 mm³ (mean = 143, SD = 105); tumor volumes measured from fDOT after Sentidye® uptake ranged from 83–265 mm³ (mean = 170, SD = 86). There was no correlation between the volumes measured with the two modalities, i.e., the ratio of fDOT-based to PET-based tumor volumes varied over almost one order of magnitude (range: 0.35–3.1; mean = 1.6; SD = 1.0). More interestingly, co-registration of fDOT with PET showed that the coefficient of overlap between vascular accumulation of Sentidye®, and tumor uptake of FDG was 42 ± 14%, indicating that only part of the tumor was hypervascular while the majority of the fDOT signal appeared located in the vasculature surrounding the tumor (Figure 2D–F).