2.1.

Experimental Setup

The experimental setup used in this study is presented in Fig. 1. The implementation of the multidetector translate-rotate scanner is similar to the one described in detail in Ref. 15. To accommodate, however, the high-frequency detectors, a novel data acquisition system was developed and higher-precision motorized scanning stages were selected as described in the following.

Fig. 1

Experimental setup. (a) Schematic drawing of the optoacoustic translate-rotate scanner showing the arrangement of the motorized stages. The linear array of detectors is mounted on the translation stage. A Cartesian coordinate system, fixed to the sample, is specified. The origin of the system is set so that the $z$ axis corresponds to the rotation axis of the rotation stage and the plane $z=0$ to the middle of the length of the linear array. The $z$ axis defines the elevation direction of the system. (b) Annotated picture of the experimental setup. (c) Picture of an excised mouse kidney embedded in a turbid agar gel. The upper-front part of the agar gel was removed to disclose the kidney, and a metric steel ruler was placed on the side for size reference.

Ultrasound detection was performed with a high-frequency 128-element linear array (LA-28.0, Vermon, Tour, France). The elements of the array had an average center frequency of 24 MHz and an average $-6\mathrm{dB}$ two-way (pulse-echo) bandwidth of 60%. The width of one element was 55 μm, with a kerf of 15 μm (i.e., a pitch of 70 μm), and its height (transverse dimension) was 1.5 mm. The elements were cylindrically focused using an acoustic lens. The focal length was $\sim 7.5\mathrm{mm}$. The geometrical characteristics of the cylindrically focused elements result in a sensitivity field with calculated full width half maximum (FWHM) dimensions in the $xy$-plane and at 24 MHz of $\sim 300\mu \text{m}$ around the focus in the transverse dimension, and longer than 1 cm for the depth-of-field.¹⁶ Detected signals were digitized at $125\mathrm{MS}/\mathrm{s}$ and with a 12-bit resolution over a 16 mV range with a 128-channel custom-built data acquisition system.

The linear array was mounted on two motorized precision stages: a translation stage (M-605.2DD, Physik Instrumente, Karlsruhe, Germany) and a rotation stage (M-062.PD, Physik Instrumente). The stages were equipped respectively with an integrated linear encoder and an integrated rotary encoder, which enabled a reliable assessment of the successive positions of the detector array during tomographic measurements of high-frequency ultrasound field. The encoders provided a micrometric accuracy for both step-by-step motion and readout of the motor positions during continuous motion. The stages were suspended from a platform so that the translation stage was rotated by the rotation stage [Fig. 1(a)]. The linear array was mounted on the translation stage so that its length axis (i.e., the long axis of the active surface of the array, parallel to the focal line of the elements) was parallel to the rotation axis of the rotation stage ($z$ axis in Fig. 1), and its median plane was perpendicular to the translation axis of the stage [Figs. 1 and 2(a)]. The minimal distance between the length axis of the array and the rotation axis was set to $R=7.5\mathrm{mm}$, i.e., the focal length of the array [Fig. 2(a)]. The data acquisition system and the motors were computer-controlled.

Fig. 2

Scanning procedures. (a) Schematic top view of the experimental setup. Schematic description of the scanning procedures in (b) the averaged acquisition mode, shown here for the first six positions of the array assuming a travel range of four to five positions [complementary patterns (15)] and (c) the continuous acquisition mode, shown here for the first five positions of the array. The positions of the array are indexed with increasing numbers. The diagrams (b) and (c), in particular the motion steps, are not to scale, but were exaggerated to illustrate better the motion of the array.

The excitation light originated from a tunable (690 to 900 nm) optical parametric oscillator laser (Phocus II, Opotek Inc., Carlsbad, California), delivering $<10\mathrm{ns}$ duration pulses with a pulse repetition frequency (PRF) of 10 Hz. The beam was guided into a silica fused-end fiber bundle (CeramOptec GmbH, Bonn, Germany) consisting of 640 fibers partitioned into four legs. To spread out light and illuminate the surface of the imaged samples more evenly, all samples were embedded in turbid gel. Turbid gel mimicking the speed of sound in soft tissues was prepared by mixing 1.6% w/m agar gel (Agar for microbiology, Fluka analytical) with 0.8% v/v intralipid-20% (Sigma). The gel was poured around the sample in a cylindrical mold of 12 mm inner diameter and 3.5 cm height so as to form an optically diffusing layer—at least 2 mm thick on the side and 5 mm thick on the top [Fig. 1(c)]. The unmolded cylinder was then inserted in a fixed sample holder so that its axis corresponded to the rotation axis of the stage and the embedded sample faced the ultrasound array. Three legs of the fiber bundle were positioned 2 cm away from the cylinder to create an illumination pattern of $\sim 10\mathrm{mm}$ height and $\sim 12\mathrm{mm}$ width on one side (cylinder surface with $y>0$), and the fourth leg was used to illuminate the top of the cylinder. The illumination was kept fixed for all the samples, and the optical excitation was performed for all the experiments at a single wavelength of 760 nm and with a per-pulse energy at the laser output of $\sim 70\mathrm{mJ}$. This wavelength ensures a good penetration in biological tissue.⁹ The laser was run continuously at 10 Hz and triggered the acquisition system.

The fiber bundle ends, the ultrasound linear array, and the sample were immersed in a water tank for acoustic coupling. The tank was filled with isotonic saline solution (0.90% w/v of NaCl in deionized water) to preserve the ex vivo sample. The solution was stabilized at room temperature.

2.2.

Scanning Procedures

The optoacoustic imaging system implemented in this study performs a translate-rotate scanning motion of the detector array around the fixed sample to acquire tomographic datasets. Full datasets are obtained by translating the detector array across the sample repetitively and rotating the combined translation stage and detector array until they have been rotated by $\sim 180\mathrm{deg}$ [Fig. 2(a)]. The translation range can be adapted to the imaged volume and should match or exceed the radial extent of the volume.¹⁵ A rotation range of $\alpha =178.5\mathrm{deg}$ and a maximum translation range $L$ of 9.0 mm were used.

Two different scanning procedures were employed in this study. In the first one, the array was scanned in discrete steps. For each detector position, five time-resolved acquisitions were averaged and the result stored in the computer before advancing to the next position. Averaging was employed to increase the signal-to-noise ratio by reducing electronic noise. The motion sequence alternated between multiple-step translation motions across the sample and single-step increments of the rotation motion [Fig. 2(b)]. This systematic scanning procedure for optoacoustic imaging was described in detail in our previous study¹⁵ and is referred to in this study as averaged acquisition mode. The rotary stage was moved here to a total of 120 positions, with a 1.5 deg angular step (${\theta }_a$). Between each rotation step, the array was scanned linearly by moving the translation stage in steps of $l_a=300\mu \text{m}$. The array trajectory was then tangent to the circle of radius $R$ centered on the axis of rotation and the positions were distributed symmetrically with respect to the point of tangency [Fig. 2(b)]. For the translation, the sampling requirement is determined by the FWHM of the sensitivity field at the center frequency of the array. The Nyquist sampling criterion implies that the translation step should be at most half of the FWHM of the sensitivity field. The step length $l_a$ corresponds to the FWHM of the sensitivity field at 24 MHz calculated for the fixed cylindrical focus of the array.¹⁶ However, as in our previous study, for two successive translations across the sample, the detector positions along the translation axis were shifted by half the step length to obtain complementary patterns. This scanning feature was implemented to obtain datasets with richer information. For the rotation motion, the angular step from our previous study¹⁵ was also used here as it showed good performance. Assessing the angular sampling requirement in terms of the Nyquist sampling criterion is actually more difficult than for the translation motion because the directivity of both the detector and the structure inside the sample have to be considered (full view criterion). No a priori assumption was made on the directivity of the absorbing structures. The angular aperture of the array in the transverse direction was $\sim 11\mathrm{deg}$. If we consider translations across the sample by pairs, the angular step was actually 3 deg, which is a reasonable oversampling in regards to the angular aperture of the array.

In the second scanning procedure, the array was continuously and simultaneously rotated and translated [Fig. 2(c)]. For each laser pulse, signals were acquired with the linear array of detectors and stored in the computer without averaging. Readouts of the motor positions given by the motion controllers were stored as well. The rotation stage was moved at constant speed in only one direction to cover the rotation range in the total duration of the scan. The translation stage operated in oscillatory motion, moving back and forth with a predefined speed $v_c$ over the same linear segment. The linear segment was chosen symmetric with respect to the point of tangency between the translation axis and the circle of radius $R$ centered on the axis of rotation [Fig. 2(a)]. For a given translation range $L$ and PRF of the laser, the translation speed $v_c$ and the number of single-pulse acquisitions per period of oscillatory motion were set so that the travel distance of the translation stage between two successive laser pulses $l_c=v_c/\mathrm{PRF}$ was approximately a predefined common fraction of $l_a$, the step length in the averaged acquisition mode (i.e., 300 μm). The total number of single-pulse acquisitions was then set so as to have 60 periods of oscillatory motion of the translation stage during the scan, and the rotation speed was determined. The number of periods of oscillatory motion was chosen equal to the averaged mode for better comparison. This scanning procedure is named continuous acquisition mode (CAM) in this study. To be able to refer easily to continuous acquisitions of different parameters, a number was associated with each CAM implementation. This number corresponds to the ratio of $l_a$ to $l_c$, the targeted travel distance of the translation stage between two successive single-pulse acquisitions in the continuous mode. The CAM number was set $\ge 1$ to ensure a good spatial sampling. Predetermined and measured parameters of continuous acquisition modes implemented for a translation range of $L=9.0\mathrm{mm}$ are given in Table 1. The observed median travel of the translation stage between successive acquisitions was lower than expected most probably because of the acceleration and deceleration phase while changing direction. Considering the maximum speed of the array during the acquisitions, the Doppler effect induced by the motion of the detectors could be considered as negligible for the measured ultrasound frequencies.¹⁷ The maximum propagation distance for the ultrasound waves being 15 mm, the displacement of the focal spot of the array during the optoacoustic wave propagation could also be considered as negligible.

Table 1

Data acquisition parameters for different modes and a translation range of 9.0 mm.

CAM #5 and the averaged acquisition mode required collecting the same number of nonaveraged signals. However, in continuous acquisition mode, the single-pulse acquisitions corresponded to successive laser pulses, while in averaged mode, each averaged acquisition started only when the motor was in position. The laser pulses occurring during the detector motion were therefore not used in averaged acquisition mode, which resulted in longer scan times (Table 1).

2.3.

Image Reconstruction and Postprocessing

Datasets acquired with the scanner were reconstructed on a grid composed of cuboid voxels of $12\mu \text{m}\times 12\mu \text{m}\times 24\mu \text{m}$, with a filtered backprojection algorithm described in detail in Ref. 15. The elevation aperture along the length axis of the array was kept constant and equal to 0.8 as long as possible. Signals were bandpass filtered (Butterworth order 3, cutoff frequencies 2 and 40 MHz).

As the translation range was shown, with the translate-rotate geometry, to define the diameter of a cylinder where image quality in terms of resolution is optimal,¹⁵ the cuboid 3-D images were cropped to remove the outer parts of the cylinder before visualization. The 3-D images were visualized using maximum amplitude projections (MAP) and a rotation around one axis. This visualization was performed with the 3-D project option of ImageJ.

2.4.

Sample Preparation

As mentioned in Sec. 2.1, all samples were embedded in turbid agar gel and molded to form cylinders of 12-mm diameter.

3.1.

Volumetric Imaging in Averaged Acquisition Mode

3.1.1.

Spatial fidelity and separation between two objects

Figure 3 displays the reconstructions of Phantom 1 and Phantom 2, composed of 20-μm diameter suture threads arranged in a cross and a loose two-loop knot, respectively. The continuous and controlled structures were scanned in averaged acquisition mode with translation ranges of 5.1 mm. The complex shape of the absorbers could be reconstructed in a volume about the height of the array [Figs. 3(a) and 3(d)] with a consistent resolution over the volume [Figs. 3(a) and 3(b)] and the multiple orientations [Figs. 3(d) and 3(e)]. One can also notice that dust particles trapped in the solid gel were reconstructed, including a large and strongly absorbing one in the second phantom around $z\sim 3\mathrm{mm}$ [Figs. 3(d) and 3(e)]. A slice orthogonal to the $z$ axis was selected from the reconstruction of Phantom 1 [Fig. 3(c)] close to the location of the intersection of the threads. This slice corresponds to the slice where the saddle point between the individual reconstructions of the threads first develops. The distance between the two maxima corresponds to the resolution in the sense of the Sparrow resolution criterion. It was determined here to be 51 μm. A second slice was taken from the reconstruction of Phantom 2 at the location of an in-plane (orthogonal to the $z$ axis) intersection of the thread [Fig. 3(f)]. This slice shows cross-sections of portions of the thread with different orientations, reconstructed with the measured tomographic dataset. Figure 3(f) reveals that objects with an elongated in-plane cross-sectional area lead to strong arc-shaped streaking artifacts in the background. For instance, artifacts in the part of the image with $x<-1\mathrm{mm}$ and $y>-1\mathrm{mm}$ are due to the crossing portions of the thread. These arc-shaped artifacts are produced by the simple radial backprojection algorithm used here for reconstruction²⁰ and the finite number of angular steps of the scan geometry. The backprojection artifacts appear weaker for objects with an isotropic in-plane cross-sectional area like the portions of the thread mainly oriented toward the $z$ axis, which can be explained by the fact that the translate-rotate scanning geometry averages spatially the artifacts produced by such objects over more projections distributed on a larger solid angle of detection. The streak artifacts are not visible on the MAP images as they have lower amplitude than the main lobes. Negative and nonphysical values also appear on the slices [Figs. 3(c) and 3(f)], and are also filtered backprojection artifacts.

Fig. 3

Reconstructions of Phantom 1 [(a) to (c)] and Phantom 2 [(d) to (f)], both composed of Ø 20 μm threads arranged respectively in a cross and a loose two-loop knot (Video 1). (a) and (b) Maximum amplitude projection (MAP) images respectively along the $x$ axis and the $z$ axis of the first phantom. (c) Two-dimensional slice corresponding to the plane marked with superimposed marks on the sides of (a). The star markers indicate the position of the maxima corresponding to the reconstruction of each thread. The distance between the two stars is 51 μm. (d) and (e) MAP images respectively along the $x$ axis and the $z$ axis of the second phantom. (f) Slice corresponding to the plane marked with superimposed marks on the sides of (d). The images are normalized by the maximum value of the three-dimensional (3-D) images. (Video 1, QuickTime, 2.1 MB) [URL: http://dx.doi.org/10.1117/1.JBO.18.10.106005.1].

3.1.2.

Resolution of isotropic objects

The third phantom, composed of Ø 10-μm microspheres randomly spread in the turbid agar gel, was imaged with the system using a translation range of 9.0 mm. Figures 4(a) and 4(b) show MAP images of the reconstruction of a 2.1-mm-thick (along the $z$ axis) slice of Phantom 3. The slice thickness was chosen such that microspheres are spread all over the diameter of the imaged volume but remain distinct and well separated in the MAP images. Microspheres at different in-plane [Fig. 4(a)] and elevation [Fig. 4(b)] positions were reconstructed. Besides amplitude variations due to light scattering, and a thresholding of the brighter voxels, all microspheres were reconstructed with a vertical rice-grain shape and 3-D resolutions consistent over the volume of the slice.

Fig. 4

Reconstruction of a 2.1-mm-thick volumetric slice of Phantom 3 (comprising Ø 10 μm microspheres randomly spread over the volume). (a) MAP image along the $z$ axis. (b) MAP image along the $y$ axis. The superimposed rectangles indicate the limits of three subregions of interest. (c), (d), and (e) MAP images of these subregions along the $z$ axis (left) and the $y$ axis (right). The amplitude of the images is normalized by the maximum value of the 3-D image. (a) and (b) are thresholds for better visualization.

The 3-D resolution of three microspheres located at different polar (along the $x$ and $y$ axes) and elevation coordinates was analyzed. Cuboid volumes that centered the microspheres ($9\times 9\times 11$ voxels) were extracted from the 3-D image [superimposed rectangles, Figs. 4(a) and 4(b)]. The MAP of these cuboids in two orthogonal planes are displayed in Figs. 4(c) to 4(e). The MAP images along the $z$ axis show that microspheres are resolved identically along the $x$ and $y$ directions. The FWHM dimensions of the MAP profiles for the three microspheres were determined to be equal to ($\text{mean}\pm \text{standard deviation}$) $29\pm 2\mu \text{m}$ in plane (along the $y$ axis) and $110\pm 5\mu \text{m}$ in elevation. By taking into account convolution effects between the finite size of the microsphere and the impulse response of the system, this measurement points to an in-plane system resolution of $<30\mu \text{m}$. The discrepancy between the two resolutions is due to the angular coverage difference induced by the limited aperture of the array along the $z$ axis.

The in-plane FWHM dimensions of the microspheres are smaller than the in-plane resolution determined with Phantom 1. Although the microspheres have a smaller diameter than the threads, both types of absorbers are expected to emit optoacoustic wave with a peak frequency greater than the highest recorded frequency.^18,19 The difference in in-plane resolution is then most probably due to the coupling between in-plane and elevation resolution for the elongated absorbing threads intersecting the $xy$ plane with an angle.

3.1.3.

Ex vivo anatomic images of mouse kidney

An excised kidney of an adult mouse was imaged ex vivo using a translation range of 9.0 mm. The MAP image along the $x$ axis is presented in Fig. 5(a). The bean-shaped structure of the kidney was positioned vertically, with the convex surface toward the illumination side. The superior part of the kidney was imaged. The renal hilum is located around $z\sim -2.4\mathrm{mm}$. Structures of the renal vascular anatomy can be visualized with MAP [Fig. 5(a) and Video 2], in particular by decreasing size and from the renal hilum to the renal cortex, dividing segmental, interlobar, and interlobular vessels. The volumetric vascular morphology visible with the obtained MAP images show similar structure as images obtained from vascular corrosion casts of mouse kidneys.²¹ Because of the inhomogeneity of the light fluence with depth, only the portion with the stronger illumination is seen on the MAP images. Correction schemes for fluence have been suggested in the literature^22,23 and will be considered as a next step of the mesoscopic setup development.

Fig. 5

Reconstructions of the kidney of an adult mouse. (a) MAP images along the $x$ axis. The superimposed marks on the sides of (a) indicate the positions of two in-plane slices. The slices are presented ordered by increasing $z$: (b) slice#1 and (d) slice#2. The images are normalized with the maximum voxel value of the 3-D image and threshold for better visualization. (c) and (e) Red-filtered photographs of the frozen and sectioned kidney corresponding respectively to slice #1 and slice #2. Legend: 1, interlobular vessels; 2, interlobar vessels; 3, segmental veins; 4, renal vein; 5, renal artery; 6, renal pelvis; 7, segmental artery; 8, medullary region with the vasa recta. The anatomical features were correlated with published anatomy.²⁴ (Video 2, QuickTime, 3.5 MB) [URL: http://dx.doi.org/10.1117/1.JBO.18.10.106005.2].

Two different slices of the 3-D image, taken around the renal hilum, are displayed in Figs. 5(b) and 5(d) and show that vessels could be imaged over the entire organ, reaching up to 10 mm away from the illuminated surface ($y<0$). Photographs of the frozen and sectioned organ for these corresponding slices are presented in Figs. 5(c) and 5(e). Black areas on these photographs are either due to blood vessels or the empty renal pelvis. As they correspond to the same kidney, the photographs validate the optoacoustic images, even if mostly large blood vessels can be correlated due to the intrinsic lack of contrast of the photographs. The renal pelvis being nonabsorbing, it could not be directly imaged with the optoacoustic system. The vascular anatomical features were also correlated with published anatomy.²⁴

The kidney was used here as a biological sample for the multiple size and branching vasculature of this organ. Despite the limited bandwidth of the array that induces a sensitivity dependence on the object size, the reconstruction demonstrated that the developed optoacoustic system was able to resolve at least five orders of vessel branching, from the renal vein to interlobular vessels. Many microvessels—arcuate as well as interlobular arteries and veins—could be resolved in the renal cortex [Video 2, Figs. 5(b) and 5(d)]. The filament-like structure of the vasa recta in the medullary region could even be visualized [Fig. 5(d)] while composed of capillaries.²¹

3.2.

Continuous versus Averaged Acquisition Mode

The mouse kidney imaged in averaged acquisition mode (Fig. 5) was scanned successively in the CAMs, listed in Table 1. Reconstructions of the two slices (slices #1 and # 2) shown in Fig. 5 for the averaged acquisition mode are presented in Fig. 6 for CAM #1, #2.5, and #5. The number of measurement positions in the averaged acquisition mode and in CAM #1 was equal (Table 1). The number of single-pulse acquisitions in the averaged acquisition mode and in CAM #5 was equal. CAM #2.5 had an intermediate number of single-pulse acquisitions.

Fig. 6

Reconstructions of slices #1 (top row) and #2 (bottom row) of the mouse kidney for several acquisition modes. The full arrows indicate regions where image quality is affected by arc-shaped streaking artifacts in average acquisition mode. The empty arrow points at interlobular vessels; the visibility of these vessels depends on the acquisition mode. The superimposed square indicates the limits of the regions for which the standard deviation of the background noise was assessed.

The image quality for the visualization of the renal artery and renal vein in slice #1, and the medullary region with the vasa recta in slice #2 increased significantly between the averaged acquisition mode and CAM #5 while the same number of single-pulse acquisitions was used. As for the knot (Fig. 3), strong arc-shaped streaking artifacts are visible for the kidney in the reconstructions of slices #1 and #2 corresponding to the averaged acquisition mode (full arrows Fig. 6). These artifacts impair the image contrast and block the visualization of fine details by increasing locally the background noise. This type of streak artifacts mostly arises from structures elongated in-plane. Figure 6 shows a streak artifact reduction with the increased CAM number, while the same filtered backprojection algorithm was used. The artifacts in CAM #1 were similar but already lower than in average acquisition mode. Therefore, this type of streak artifacts occurs here if too few views of the sample were obtained. Continuous acquisitions increased the number of views of the sample since the array was constantly moving while keeping the number of single-pulse acquisitions inferior or equal to the average mode.

In areas free of strong artifacts, some interlobular vessels clearly visible in averaged acquisition mode (empty arrows Fig. 6) could not be visualized in CAM #1 and hardly in CAM #2.5. Two factors could be responsible. First, as reconstructions for CAM #1 and #2.5 use datasets collected with less single-pulse acquisitions than in averaged acquisition mode, electronic noise reduction by averaging is weaker and therefore the background noise of the image could be stronger. Second, as in continuous acquisition mode, the angular position of the rotation stage is different for each measurement; the number of views for a directional absorbing structure could be lower than with a systematic scanning across the sample for each angular step. To test the first hypothesis, the standard deviation of background noise was computed in the same cuboid volume of $21\times 21\times 21$ (i.e., 9261) voxels (around slice #2 and marked with a superimposed square in Fig. 6) for all the acquisition modes (Table 1). The results are displayed in Fig. 7 and show that the background noise decreases with the CAM number. In comparison to the averaged acquisition mode, the background noise was found 65% stronger in CAM #1, 31% weaker in CAM #5, and only 5% stronger in CAM #2.5. These results indicate that the noise level could be a dominant factor to explain the nonappearance of some interlobular vessels in CAM #1. The background noise was found very similar for the averaged acquisition mode and CAM #2.5 while half the number of single-pulse acquisition was used for the latter, and lower for CAM number superior or equal to 3. The background noise is therefore mostly due to the streak artifacts in averaged mode, and to electronic noise for CAMs with numbers $<2.5$. For CAM #2.5, the limited number of views could explain the contrast difference for the spotted interlobular vessels.

Fig. 7

Standard deviation of the background noise as a function of the number of continuous data acquisition modes for the volumetric image of the kidney. The standard deviation was computed in $21\times 21\times 21$ voxel cuboids. Each volumetric image was normalized by the number of measurement positions. The horizontal dashed line corresponds to the value in averaged data acquisition mode.

Phantom 3, composed of Ø 10-μm black microspheres (Fig. 4), was also imaged in continuous modes with numbers ranging from 1 to 5. Besides an increase of the background noise for CAM with numbers $<3$, the 3-D resolution of the three microspheres analyzed in Sec. 3.1.2 (Fig. 4) was found unchanged between averaged or continuous modes.

In brief, continuous acquisition modes have the advantages of reducing the scan time, the backprojection streaking artifacts, and the background noise, without any drawbacks except for the large size of the tomographic datasets generated, and therefore the computationally demanding reconstructions.