Open Access
Add to BibSonomy
Abstract
Photoacoustic imaging has demonstrated recent promise for surgical guidance, enabling visualization of tool tips during surgical and non-surgical interventions. To receive photoacoustic signals, most conventional transducers are rigid, while a flexible array is able to deform and provide complete contact on surfaces with different geometries. In this work, we present photoacoustic images acquired with a flexible array transducer in multiple concave shapes in phantom and ex vivo bovine liver experiments targeted toward interventional photoacoustic applications. We validate our image reconstruction equations for known sensor geometries with simulated data, and we provide empirical elevation field-of-view, target position, and image quality measurements. The elevation field-of-view was 6.08 mm at a depth of 4 cm and greater than 13 mm at a depth of 5 cm. The target depth agreement with ground truth ranged 98.35-99.69%. The mean lateral and axial target sizes when imaging 600 μm-core-diameter optical fibers inserted within the phantoms ranged 0.98-2.14 mm and 1.61-2.24 mm, respectively. The mean ± one standard deviation of lateral and axial target sizes when surrounded by liver tissue were 1.80±0.48 mm and 2.17±0.24 mm, respectively. Contrast, signal-to-noise, and generalized contrast-to-noise ratios ranged 6.92–24.42 dB, 46.50–67.51 dB, and 0.76–1, respectively, within the elevational field-of-view. Results establish the feasibility of implementing photoacoustic-guided surgery with a flexible array transducer.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Photoacoustic imaging has recently demonstrated strong viability for surgical guidance, enabling visualization of instrument tips during surgical and non-surgical interventions [1–3]. In photoacoustic imaging, a target of interest illuminated by a light source (e.g., a pulsed laser) absorbs the light, causing thermal expansion and the production of acoustic waves that propagate toward the surface of the tissue. These acoustic waves can be received by an external ultrasound transducer to create images after implementing appropriate time delays [1,4,5]. The high spatial resolution, strong optical absorption contrast, and centimeter penetration depth of photoacoustic imaging make it an excellent technique for surgical guidance [2,6]. However, many surgical resection plans (e.g., to remove hyperfunctioning parathyroid glands [7] or pancreatic tumors [8]) rely on pre-operative medical images, such as positron emission tomography (PET), x-ray computed tomography (CT), or magnetic resonance imaging (MRI), where real-time visualization of surgical tools is missing. In such cases, accidental injuries to critical internal structures are possible [1]. To avoid accidental injury, one or more optical fibers may be inserted within or externally appended to the tips of surgical tools or instruments (e.g., needle tips [9–11], catheter tips [12], drill tips [13,14]). This arrangement enables visualization of tool tips relative to surrounding regions of interest, as signals from both the fiber tip and the tool tip, as well as the surrounding region of interest, can be simultaneously received by the transducer [1,2].
Conventional linear array transducers (e.g., clinical hand-held probes, laparoscopic probes) have rigid sensing parts with fixed array geometries, making them ideal for flat surfaces. When placed on uneven surfaces, to avoid the degradation of image quality caused by air gaps between the transducer and tissue surface [15], either a large amount of ultrasound gel is needed to fill the gap or additional contact pressure is required. This pressure can cause organ distortions, tool tip localization difficulties, possible patient discomfort during interventions that lack anesthesia, and risk of further injury to tissue [16]. The development of advanced sensing technology, including flexible array transducers, has recently opened new opportunities for photoacoustic-guided surgery applications. Instead of the rigid geometry of traditional probes, a flexible array is able to deform and provide complete contact on body parts with different curved surfaces [17], including the skull, spine, and abdomen, which is particularly beneficial for photoacoustic-guided neurosurgery [18], spinal fusion surgery [19,20], liver surgery [21], and hysterectomy [22]. This type of complete contact minimizes the amount of required ultrasound gel, organ deformations, anatomical distortions, and associated patient discomfort. In addition, due to the minimized deformations, distortions, and gel requirements, complete contact is expected to improve overall target visualization and localization.
Major challenges with utilizing the flexible array transducer include array geometry estimation and image reconstruction. To measure transducer element positions, a shape sensing fiber can be attached along the array [23]. Different shape estimation algorithms have been proposed for ultrasound image formation, such as updating shape parameters based on image sharpness evaluation [24] or image entropy calculation [25] or tracking element positions in real-time using optical trackers [26]. For unknown array shapes, deep neural networks were demonstrated as alternatives to output array geometry parameters [27] or to directly generate ultrasound images without distortions [28]. Spinal deformity was also measured when employing a flexible array ultrasound imaging transducer in a phantom study with four known shapes (i.e., flat, convex, concave, and s-shaped) [29]. With respect to photoacoustic signal reception, a flexible Lead Zirconate Titanate (PZT) transducer was fabricated and demonstrated to receive A-line signals [30]. In addition, photoacoustic images with custom-designed flexible transparent capacitive micromachined ultrasound transducers (CMUTs) were created and organized in a curved geometry for through-illumination systems [31].
Our previous conference papers introduced the first known demonstration of photoacoustic imaging with a flexible array placed on curved surfaces for potential interventional applications [32], with performance compared to that of a laparoscopic probe [33]. In addition, we derived time-delay calculations for photoacoustic image formation with the flexible array [32]. This paper builds on our previous conference papers [32,33], with four new contributions to complement our initial demonstrations of feasibility. First, our mathematical equations to calculate element positions of the flexible array in known concave shapes were added and validated with simulations. Second, lateral location measurements for multi-target phantom experiments were added to supplement previous axial location measurements, thus providing a more complete target localization assessment, which is important for localization characterization of 2D images. Third, in addition to the four hemispherical plastisol phantoms we previously constructed, including three single-target phantoms with different radii of curvature and one multi-target phantom, we introduce a fifth ex vivo bovine liver phantom and present photoacoustic images and associated analyses obtained with the flexible array transducer. Finally, we present the first known co-registered photoacoustic and ultrasound images acquired with a flexible array transducer to enhance the level of information provided by either imaging modality independently. This combination of mathematical theory and validation for image formation, associated analyses with single and multiple image targets, and ex vivo validation provides a comprehensive demonstration of the promising application of flexible array transducers to visualize tool tips in photoacoustic-guided surgeries or interventions.
The remainder of this article is organized as follows. Section 2 describes the image formation theory and associated assumptions, simulation validation methods, experimental components utilized and constructed, the phantom and ex vivo experiment procedures, and our image analysis methods. Section 3 presents the photoacoustic images and assessments with the flexible array transducer. Section 4 discusses our findings and the associated implications for future research. Section 5 summarizes the major contributions of this paper.
2. Methods
2.1 Image reconstruction theory
To create images with a flexible array transducer, we assume complete contact between the transducer and a curved surface. Thus, the geometry of the flexible array transducer is derived based on a given radius of curvature. This is a reasonable assumption if a transducer is fixed on a curved body part and measurements of the curvature can be obtained prior to photoacoustic imaging. Measurements can also be obtained with optical trackers [26,34].
Unlike a linear array placed on a flat surface where all the element positions are fixed, the geometry of the flexible array changes according to the surface, based on the following definition:
where $\theta _i$ is any value in the range $\theta _1$ to $\theta _N$ incremented by integer (i.e., $\mathbb {Z}$) multiples of $\theta _\Delta$, which is the angle between adjacent elements, and $N$ denotes the total number of elements in the transducer. With the origin of a polar coordinate system centered within the 1D array, resulting in the relationship $\theta _1=-\theta _N$, the angle between the first and last elements is defined as:With a known element pitch of arclength $P_\Delta$ and a radius of curvature $R$, the angular increment between elements of the flexible array transducer (i.e., $\theta _\Delta$) can be calculated as follows:
Using Eqs. (2) and (3), the angle $\theta _i$ between the center of the transducer and the $i$th element of the transducer can be computed as:
To relate the angles described above to a Cartesian coordinate system with an origin located at the center of the transducer, the $x$, $y$, and $z$ positions of the $i$th element can be written as:
Therefore, Eqs. (1)–(5) define the conversion between Cartesian and polar spatial coordinates when describing the shape of a flexible array.
Photoacoustic images may then be reconstructed using conventional delay-and-sum (DAS) beamforming. The time delay, $\tau _i$, for each transducer element in dynamic receive DAS beamforming with the flexible array is described as follows [32,33]:
Fig. 1. Illustrations of (a) key beamforming parameters and (b) the scan conversion process (i.e., from matrix data to a sector shape) for image display. The red line indicates a scan line, and the black dot associated with ($x_f,z_f$) represents a dynamic focal point.
To form the final image, the DAS-beamformed signals are envelope detected, scan converted, normalized to the brightest pixel, and log compressed. Scan conversion (graphically illustrated in Fig. 1(b)) uses bilinear interpolation [35] to transform a 2D matrix of data with coordinates $(r_{n},\theta _{l})$ to Cartesian coordinates $(x,z)$, where the reconstructed photoacoustic image is displayed in a sector shape when placing the flexible array on a curved surface. These same equations and transformations were implemented to achieve co-registered ultrasound images after accounting for the ultrasound transmission sound travel and associated time delays.
2.2 Simulation setup
To validate the image reconstruction theory presented in Section 2.1 and determine the impact of incorrect array geometry estimation, simulations were performed using the k-Wave toolbox [36]. The flexible array transducer was simulated with 128 elements, 1 mm element pitch, and 5 MHz center frequency. A diagram illustrating various simulation setups with different assumed transducer array shapes is shown in Fig. 2(a). The ground truth surface curvature $R_\text {true}$ is 81.3 mm. The two additional curvatures shown in blue and green in Fig. 2(a) represent examples of wrong radii implemented for DAS beamforming (i.e., $R_\text {DAS}\approx R_\text {true} \pm 10 mm$). The location of the transducer center was fixed for each curvature investigated. In addition, an irregular surface was modeled with a curvature of 81.3 mm and a sinusoidal shift. The shift was applied to the 45$^\text {th}$ through 84$^\text {th}$ elements, as illustrated in Fig. 2(a). A single round acoustic target was modeled at a depth of 40 mm from the center of the transducer with a diameter of 0.6 mm or 2 mm. The sound speed surrounding the target was set to 1540 m/s.
Fig. 2. (a) Regular (concave) and irregular array shapes with the round acoustic target at the ground truth depth of $D_g$. (b) Time delay errors as a function of radii used in delay-and-sum (DAS) beamforming (i.e., $R_\text {DAS}$), with the left and right vertical dashed lines indicating $R_\text {DAS}$ values of 71 mm and 91 mm, respectively.
To quantify the impact of incorrect array shapes on the overall image reconstruction process, a shape estimation error, converted to a time delay error, $\epsilon _\tau$, through the known sound speed, was computed as the root-mean-squared-error (RMSE) of the time delay differences for matched elements in the correct (i.e., subscript $i\_\text {true}$) and incorrect (i.e., subscript $i\_\text {DAS}$) array shapes:
2.3 Experimental setup
Our photoacoustic imaging system consisted of a Phocus Mobile laser (OPOTEK, Carlsbad, CA, USA) connected to a Vantage 128 ultrasound system (Verasonics, Kirkland, WA, USA). The laser was operated at a wavelength of 750 nm with a pulse width of 5 ns and a pulse repetition rate of 10 Hz. There were two optical delivery methods: (1) a 600 $\mu$m core-diameter silica optical fiber and (2) a 1-to-7 fan-out optical fiber bundle (Thorlabs, Newton, NJ, USA) comprised of individual 600 $\mu$m core-diameter silica optical fibers. Each optical delivery option was independently connected to the laser system based on the experimental setup and goals. The average laser energy output at each individual optical fiber tip was measured with an external energy meter (Ophir, North Logan, UT, USA) prior to each experiment. For the single optical fiber, the energy was 550 $\mu$J per pulse. For three fibers of the 1-to-7 optical fiber bundle, each emitted average energies of 257 $\mu$J, 335 $\mu$J, or 348 $\mu$J per pulse. Emissions from the laser were synchronized with the ultrasound receiving electronics. A JAR1109 flexible array transducer (Japan Probe and Hitachi, Japan) was connected to the Vantage ultrasound system, and raw ultrasound and photoacoustic data were acquired sequentially for image registration. The raw ultrasound image data were obtained after a single 0$^\circ$ plane wave transmission sequence. The parameters of the flexible array transducer are listed in Table 1.
Table 1. Flexible array transducer parameters
Fig. 3. Cross-sectional schematic diagrams and photographs of the experimental setups implemented to image the (a,c) single- and (b,d) multi-target phantoms. The red circles denote the optical fiber tips. $D_g$ is the ground truth depth of the single target and $R$ is the radius of curvature of the phantom. The numbers in (b) represent the ground truth depths and lateral positions in units of $mm$.
To obtain photoacoustic images with the flexible array placed on surfaces of various curvatures, three hemispherical phantoms with 81.3 mm, 63.6 mm, and 50.8 mm radii of curvature (hereafter referenced as the large, medium, and small phantoms, respectively) were constructed using plastisol. Although multiple possible photoacoustic phantom construction materials and designs exist [37], we chose plastisol because of its construction simplicity and its stability over time [38]. In addition, optical and acoustic scatterers were not included in the phantom to create an ideal environment for evaluating the DAS beamformer [39]. A hollow channel was created in each plastisol phantom to insert a 2 mm-diameter needle housing the individual optical fiber described above at fixed depths of 40 mm, 50 mm, and 40 mm from the peak of the curved surfaces of the large, medium, and small phantoms, respectively. The flexible array transducer was placed on the curved surface of the phantom for photoacoustic data acquisition with complete contact. A schematic diagram and photograph of this experimental setup are shown in Figs. 3(a) and 3(c), respectively. This experimental design is relevant because some interventional images may only have a single target (e.g., a needle tip), and we present this possibility for multiple array curvatures.
To investigate expectations when multiple photoacoustic targets are present, a fourth hemispherical plastisol phantom with 81.3 mm radius of curvature was constructed. Figure 3(b) shows a schematic diagram of this multi-target phantom, which contains three hollow channels at depths of 40 mm, 50 mm, and 60 mm from the peak of the curved surface. The lateral distances between the centers of the adjacent hollow channels are 10.23 mm and 8.82 mm. The diameter of each hollow channel is 5.3 mm. Figure 3(d) shows the photograph of the multi-target phantom experimental setup. Three of the seven optical fibers from the 1-to-7 fiber bundle described above were each inserted into one of the three channels described above, with greater output laser energies measured at the fiber tip placed at deeper depths. Ten photoacoustic images were acquired for each fiber location setup described for the single- and multi-target phantom experiments.
Fig. 4. (a) A 2D cross-section schematic diagram and (b,c) photographs of the experimental setup implemented to image the ex vivo bovine liver.
To further evaluate capabilities in the presence of scattering tissue, a fifth hemispherical phantom was constructed by embedding ex vivo bovine liver tissue in the previously presented phantom designs, with a schematic diagram and photograph shown in Figs. 4(a) and 4(b), respectively. The radius of curvature of the phantom surface was 81.3 mm. The top surface of the liver was placed at a depth of 27 mm from the peak of the curved surface. A 2 mm-diameter hollow needle with the 600 $\mu$m-core-diameter optical fiber described above was inserted into the liver. The flexible array transducer was placed on the curved surface of the phantom for photoacoustic data acquisition. A photograph of the ex vivo experimental setup, including the flexible array (which blocks the view of the ex vivo tissue), is shown in Fig. 4(c). The ultrasound and photoacoustic images acquired with this phantom were further validated by replacing the flexible array transducer with a linear laparoscopic probe with a center frequency of 7.5 MHz, 80% bandwidth, 128 elements, elevation aperture of 5 mm, and elevation focus of 20 mm.
The sound speed of plastisol is known to be less than 1540 m/s [37,38], and the sound speed of liver can vary [40]. However, we maintained a sound speed of 1540 m/s when reconstructing images of the five phantoms described above to maintain consistency of this parameter across all experimental results, which matches conventional approaches to image display [4].
2.4 Image quality and target position measurements
To assess image quality, target size and visibility were quantified. The size of the optical fiber tip in the scan-converted photoacoustic image was determined based on lateral and axial full-width at half-maximum (FWHM) measurements. Target visibility was determined with contrast, signal-to-noise ratio (SNR), and generalized contrast-to-noise (gCNR) measurements [41–43], calculated as follows:
where $\mu _t$ and $\mu _b$ are the means, $h_t$ and $h_b$ are the histograms, and $\sigma _b$ is the standard deviation of the signal amplitudes (after envelope detection, scan conversion, and normalization) within 0.6 mm $\times$ 0.6 mm regions of interest (ROIs) placed at the same image depth within the photoacoustic target (denoted by subscript $t$) or within the background of the photoacoustic image (denoted by subscript $b$), $N_{bin}$ = 256 histogram bins, and $x_k$ is the mean value of the $k$th bin. The target ROI was centered on the brightest pixel in the region being evaluated, and the background ROI was laterally displaced by 4.85 mm from the target ROI.To determine target positions relative to ground truth values, a 10 mm $\times$ 10 mm ROI was first selected surrounding the photoacoustic target (i.e., the optical fiber tip) in each image. The $z$ (i.e., axial) position associated with the maximum brightness within this ROI was defined as the target depth, $D_t$. To determine the target depth agreement with the ground truth (i.e., known fiber tip depth based on the phantom design), depth accuracy for multiple measurements was defined as:
where $\overline {D}_t$ is the mean target depth across multiple measurements, and $D_g$ is the ground truth depth of the center of the hollow channel in the phantom, verified with post-construction caliper measurements.In the experimental data, considering that the diameter of the hollow channel is larger than that of the optical fiber, the exact lateral ground truth of the fiber tip position is unknown. Instead, the range of the lateral distances between adjacent hollow channels (i.e., adjacent targets) was calculated as follows:
where $C$ is the distance between center locations of the hollow channels measured with caliper and $D$ is the diameter of the hollow channel.2.5 Elevation field-of-view estimation
To empirically approximate the elevation field-of-view (FOV) of the flexible array transducer, the optical fiber tip was first translated in the elevation dimension of the transducer (i.e., orthogonal to the axial and lateral dimensions) when imaging the single-target phantoms. The optical fiber was fixed to a manual translation stage (Thorlabs, Newton, NJ, USA). The flexible array was positioned to visualize the fiber tip in the photoacoustic image with the elevation axis of the probe approximately parallel to the optical fiber. The fiber was translated in increments of 1 mm, for a total travel distance of 15 mm (i.e., the maximum travel distance of the translation stage), and ten photoacoustic images were acquired at each fiber position. The maximum brightness in each of the ten acquired photoacoustic images was calculated and plotted as a function of elevation fiber positions. The average of maximum brightness values for each of the fiber positions was plotted as a brightness curve, which was then normalized to the range [0, 1]. The FWHM was determined from this normalized brightness curve, after linearly interpolating to increase the number of samples by a factor of 100 (i.e., resulting in a precision of 0.01 mm). When the optical fiber translation process produced an incomplete brightness curve with no clear minimum, the elevation FOV was determined based on target visibility measurements, as described in our previous work [33]. In particular, the minimum contrast and SNR within the elevation FOV was used as a surrogate to determine elevation FOV when a complete brightness curve was produced with the flexible array imaging at least one of the three hemispherical phantoms.
To compensate for anticipated laser energy fluctuations, outliers of the measured contrast and SNR values at each fiber position were removed when determining the elevation FOV. As ten photoacoustic images were acquired at each position, the removal of outliers (defined as values greater than 1.5 times the interquartile range) provided results based on relatively stable laser energies.
3. Results
3.1 Simulation results
Figure 5 shows example photoacoustic images obtained with simulated data containing individual targets with diameters of 0.6 mm (top row) or 2 mm (bottom row). The ground truth images (first column) contain correct target sizes and depths. However, wrong radii (second and third columns) for DAS reconstruction severely altered the lateral target sizes with similar axial target sizes to that of the ground truth images. When $R_{\text {DAS}}$ was closest to the irregular surface shape (fourth column), the lateral and axial target sizes in the reconstructed images were both similar to those of the ground truth images. Although more sidelobes were observed in photoacoustic images reconstructed with incorrect time delays (i.e., incorrect $R_{\text {DAS}}$ and irregular shape), targets were similarly visible in all cases. Similarly, the target depths of the brightest pixels were close to the ground truth depth of 40 mm in all cases.
Fig. 5. Photoacoustic images of 10 mm $\times$ 10 mm regions surrounding simulated 0.6 mm-diameter (first row) and 2 mm-diameter (second row) targets, reconstructed with correct (first column) and incorrect (second and third columns) transducer radii. When an irregular surface was simulated, images were reconstructed with the closest possible transducer radius (fourth column). Images are displayed with 15 dB dynamic range.
Fig. 6. Plots of (a) lateral target size, (b) axial target size, (c) target depth agreement, (d) contrast, and (e) SNR as functions of $R_{\text {DAS}}$.
Figure 6 quantifies the measured axial and lateral target sizes, target depths, contrast, and SNR as functions of $R_\text {DAS}$ for the simulated data. In Fig. 6(a), the lateral target sizes are similar to the corresponding ground truth sizes when $R_\text {DAS}$ is closest to $R_\text {true}$, with greater deviations from the ground truth obtained as the difference between $R_\text {DAS}$ and $R_\text {true}$ increases. In Fig. 6(b), as the difference between $R_\text {DAS}$ and $R_\text {true}$ increased, the axial target sizes were similar to the ground truth, which agrees with our qualitative observations of Fig. 5. The relatively smaller axial target sizes farther away from the true curvature are consistent with expectations for conservation of signal energy relative to the larger lateral sizes at similar values of $R_\text {DAS}$. In Fig. 6(c), the target depth agreement differed by less than 0.6% from 100% and 98% for the 0.6 mm and 2 mm targets, respectively. In Figs. 6(d) and 6(e), contrast and SNR values were maximized when $R_\text {DAS}=R_\text {true}$ for both target sizes, and these values generally decreased as $R_\text {DAS}$ deviated from the ground truth. Similar trends were observed for the irregular surface, with the exception of depth agreement, which was less stable yet >95% as a function of $R_\text {DAS}$. Although the contrast and SNR ranges were large in each case (i.e., contrast ranged 22.27-47.54 dB and 15.16-38.37 dB and SNR ranged 64.27-91.31 dB and 57.06-82.19 dB for curved and irregular surfaces, respectively), gCNR values were 1 for all cases, demonstrating that the photoacoustic targets were detectable, as described in more detail in [41]. In each case, the DAS beamformer performed best on the irregular surface shape when assuming the closest possible fixed curvature (i.e., $R_\text {DAS}$=81.3 mm).
3.2 Single-target phantom results with multiple surface radii
Figure 7 shows an example registration of photoacoustic and ultrasound images acquired with the flexible array placed on the curved surface of the large, single-target phantom. The images are presented in a sector shape after scan conversion. The fiber tip can be clearly identified as a single target at the depth of 40 mm in the photoacoustic image. In the ultrasound image, the area with the high-amplitude signals at 40 mm depth is larger than the target observed in the photoacoustic image, which occurs because the optical fiber resides within a hollow needle with a larger diameter than the fiber. The second bright area shown in the ultrasound image at the depth of 45 mm is the remnants of the channel where the fiber-needle pair was placed during a previous insertion (which presents as a hyperechoic region due to acoustic impedance mismatches between the air and plastisol within the needle track left behind in the plastisol phantom [39]).
Fig. 7. Photoacoustic image overlaid on co-registered ultrasound image with the fiber tip at the depth of 40 mm in the large phantom.
Figure 8 shows results obtained with the flexible array imaging the medium phantom. In Fig. 8(a), photoacoustic images are presented in 10 mm $\times$ 10 mm ROIs with one example image shown for each stationary fiber position throughout the translation process. Qualitatively, the target is unclear at the position of 0 mm (i.e., when the fiber tip was located off-axis relative to the imaging plane of the transducer). The target is qualitatively visible from positions 1 mm to 8 mm, when located within the elevation FOV of the flexible array. At positions 9 mm and 10 mm, the target is no longer visible, as it was located outside of the elevation FOV. In Fig. 8(b), the distribution of the corresponding maximum brightness values in each of the 10 photoacoustic images for each of the 11 fiber positions is shown in the box-and-whisker plot. The horizontal line and circle denote the median and mean, respectively, of the brightness values in the ten images at each stationary fiber position. The top and bottom of each box indicate the upper and lower quartiles, respectively. The maximum and minimum of the data set at each position are represented by the top and bottom horizontal lines of the whiskers, excluding outliers with maximum amplitudes larger than 1.5 times the interquartile range, which are represented by the crosses. The corresponding elevation FOV of the flexible array (i.e., the FWHM of the normalized brightness curve, as defined in Section 2.5) is 6.08 mm. Within this elevation FOV, the minimum contrast and SNR values were 9.74 dB and 51.81 dB, respectively, which were used to set target visibility thresholds when using the surrogate method described in Section 2.5 to define the elevation FOV of the flexible array.
Fig. 8. (a) Photoacoustic images of the fiber tip translated along the elevation dimension of the transducer at approximately 50 mm depth in the medium phantom. Example target and background ROIs utilized to calculate contrast, SNR, and gCNR are shown in the image acquired at 0 mm as green and blue boxes, respectively. (b) The box-and-whisker plot shows the distribution of the maximum brightness values at each fiber position and the dashed line shows the normalized mean values of maximum brightness at each fiber position (i.e., the normalized brightness curve for elevation FOV measurements).
Fig. 9. Contrast and SNR measurements of the photoacoustic images acquired with the flexible array transducer placed on curved surfaces of (a) the large phantom and (b) the small phantom. The dashed line in each plot denotes the contrast or SNR threshold obtained when imaging the medium phantom.
Figure 9 shows distributions of the contrast and SNR of photoacoustic images acquired with the flexible array imaging the large and small phantoms. The transducer elevation FOVs cannot be determined by the brightness curves of the large and small phantom experiments because the curves are not complete (see Fig. 3 in [33] for an example from the large phantom), yet the photoacoustic target is qualitatively visible at all fiber positions during the translation process. Therefore, the respective 9.74 dB and 51.81 dB contrast and SNR thresholds obtained when imaging the medium phantom were implemented to determine the elevation FOVs of the flexible array in the large and small phantom experiments. In Fig. 9(a), the medians of the contrast and SNR from positions 0 to 14 mm reside above the corresponding thresholds. Therefore, the elevation FOV of the flexible array transducer is considered to be at least 14 mm at an image depth of 4 cm when imaging the large phantom. In Fig. 9(b), the medians of the contrast and SNR from positions 0 to 13 mm reside above the corresponding thresholds. Therefore, the elevation FOV of the flexible array is considered to be at least 13 mm at an image depth of 4 cm when imaging the small phantom.
Table 2 summarizes quantitative results obtained within the elevation FOV grouped by phantom size. The first row reports the elevation FOV, which varies based on target depth. The second row reports target depth agreement with the ground truth. The mean target depth within the elevation FOV was 40.33 mm, 50.16 mm, and 39.77 mm for the large, medium, and small phantoms, respectively, resulting in 99.17%, 99.69%, and 99.43% depth agreement with the ground truth, respectively. While incorrect sound speeds are known to interfere with target depth accuracy [44], the same phantom material and sound speed were used for the three phantoms, thus sound speed errors are unlikely to contribute to the minimal depth agreement differences observed across the three phantoms. In addition, the target depth agreement values for the large phantom in Table 2 are within the range of the two target sizes simulated for the same phantom surface radius in Fig. 6(c).
Table 2. Summary of quantitative results obtained with the large, medium, and small phantoms, including elevation FOV, target depth agreement with ground truth, mean$\pm$one standard deviation of axial and lateral target sizes, and contrast, SNR, and gCNR target visibility ranges
The third and fourth rows of Table 2 report the mean and standard deviation values of the lateral and axial target size measurements within the elevation FOV, revealing differences between the measured sizes and the ground truth. The lateral and axial target sizes obtained with the large phantom in Table 2 are generally within the range of the two target sizes simulated for the same phantom surface radius in Figs. 6(a) and 6(b), respectively, indicating consistency between simulated and experimental results.
The fifth and sixth rows of Table 2 summarize the target visibility measurements in terms of contrast and SNR within the elevation FOV (with outliers removed, as noted in Section 2.5). The minimum contrast and SNR values obtained with the medium phantom were 9.74 dB and 51.81 dB, respectively, which were used to set the target visibility thresholds for the large and small phantom experiments, as described when reporting elevation FOV results in the first row of Table 2. Although the contrast and SNR ranges in Table 2 do not overlap with corresponding results obtained in Figs. 6(d) and 6(e), respectively, it is known that these metrics can largely vary [41]. In such cases, gCNR has been determined to be more useful, and the consistency between simulated and experimental target visibility results for the large phantom is better supported by the gCNR range of 0.96–1 in the final row of Table 2.
3.3 Multi-target phantom results
Figure 10 shows example photoacoustic images of three targets acquired with the flexible array imaging the multi-target phantom. The mean target depths across the ten images were 40.37 mm, 50.38 mm, and 59.01 mm for the 40 mm, 50 mm, and 60 mm target depths, respectively, corresponding to 99.08%, 99.24%, and 98.35% depth agreement, respectively. The diameter of the hollow channel in the phantom is 5.3 mm. Therefore, the ranges of the lateral distance ground truths between targets 1 and 2 and targets 2 and 3 were 4.94–15.53 mm and 3.52–14.12 mm, respectively. The mean lateral distances measured between the brightest pixels within targets 1 and 2 and targets 2 and 3 were 6.42 mm and 3.59 mm, which reside within the associated ranges.
Figure 11 displays violin plots of the measured target sizes and target visibility (i.e., contrast and SNR), where the shape of the shaded colors represent the probability density of the underlying data, each solid gray box denotes the interquartile range, and each circle denotes the median. In Fig. 11(a), the median lateral and axial sizes are 1.12 mm and 1.60 mm, 1.16 mm and 1.36 mm, and 1.12 mm and 1.92 mm, for targets 1, 2, and 3, respectively. The maximum target size deviation from the ground truth is 1.56 mm, which was measured along the axial dimension with target 3 at the depth of 6 cm. The mean $\pm$ one standard deviation of combined target size measurements for the multiple targets are 1.16$\pm$0.12 mm in lateral dimension and 1.61$\pm$0.26 mm in axial dimension. In Figs. 11(b) and 11(c), the contrast and SNR range 13.87–24.42 dB and 56.83–67.51 dB, respectively, indicating good target visibility, which is also supported by the measured gCNR of 1 for the three targets [41].
Fig. 10. Photoacoustic images of (a) three targets located at approximately 40 mm, 50 mm, and 60 mm depths in the multi-target phantom and corresponding 10 mm $\times$10 mm ROIs surrounding (b) Target 1, (c) Target 2, and (d) Target 3. Example target and background ROIs for calculating contrast, SNR, and gCNR are shown as green and blue boxes, respectively.
Fig. 11. Violin plots demonstrating (a) the measured size of the three targets located in the multi-target phantom and the target visibility in terms of (b) contrast and (c) SNR.
3.4 Validation with ex vivo liver tissue and laparoscopic transducer
Figure 12 shows co-registered ultrasound and photoacoustic images of an ex vivo bovine liver containing an optical fiber tip, acquired with both laparoscopic and flexible array transducers. Ultrasound images acquired with the laparoscopic probe and the flexible array are displayed with 60 dB and 63 dB dynamic ranges, respectively. The flat linear structure with high amplitude at the depth of 27 mm in both images denotes the top surface of the liver which was in contact with the plastisol. The needle tip is difficult to be detected in both B-mode ultrasound images alone, due to the presence of the hyperechoic liver tissue structures. With the photoacoustic images overlaid, the fiber tip is clearly visualized below the liver surface. There are additional photoacoustic signals that are likely due to a combination of retained blood in the ex vivo liver tissue and scattering from the presence of tissue. These additional signals are present in both the laparoscopic and flexible array photoacoustic images, and they are not present in Figs. 7 and 10, which were acquired in the absence of ex vivo tissue. Additional comparisons to the laparoscopic transducer characterization with an associated photoacoustic image analysis in the absence of ex vivo tissue are available in our previous publication [33].
Fig. 12. Photoacoustic images with co-registered ultrasound images from the ex vivo bovine liver phantom acquired with (a) the laparoscopic transducer and (b) the flexible array transducer. There are additional photoacoustic signals that are likely due to a combination of retained blood in the ex vivo liver tissue and scattering from the presence of tissue.
Figure 13 shows violin plots of target size and target visibility measurements obtained with the flexible array and the laparoscopic transducer when imaging the ex vivo bovine liver phantom. In Fig. 13(a), the medians of the lateral and axial size measured in photoacoustic images acquired with the flexible array were 1.64 mm and 2.12 mm, respectively, compared to 0.96 mm and 1.60 mm, respectively, when images were acquired with the laparoscopic transducer. In Fig. 13(b), the contrast ranges 13.93–20.40 dB and 15.09–19.07 dB in images acquired with the flexible and laparoscopic transducers, respectively. In Fig. 13(c), the SNR ranges 57.14–63.57 dB and 53.72–57.61 dB in images acquired with the flexible and laparoscopic transducers, respectively. The median values of the contrast and SNR measured with the flexible array are larger than those of the laparoscopic transducer. However, the same ROIs produced median gCNR values of 1 in both cases, indicating that both targets are satisfactorily detectable [41]. The mean target depths across the ten photoacoustic images are 35.99 mm and 33.94 mm when imaging with the flexible and laparoscopic transducers, respectively, with differences likely due to the additional phantom deformation required for acoustic coupling with the laparoscopic probe.
Fig. 13. Violin plots of (a) target size, (b) contrast, and (c) SNR measured with photoacoustic image acquired in ex vivo bovine liver phantom using the flexible array and laparoscopic transducer.
4. Discussion
The work presented in this paper establishes the feasibility of utilizing a flexible array transducer to visualize tool tips in photoacoustic-guided surgical or interventional applications. We have comprehensively documented and demonstrated multiple important details surrounding utility and deployment for this task. Within the elevation FOV, the single-target, multi-target, and ex vivo bovine liver phantom experiments achieved target depth agreement, target sizes, and target visibility that are consistent with expectations based on corresponding simulated results (Fig. 6) and are also within previously reported ranges for photoacoustic technology [41]. The qualitative and quantitative simulation results (Figs. 5–6) additionally demonstrate the impact of incorrect array geometry estimation on image quality. In addition, co-registered photoacoustic and ultrasound images were successfully presented to provide more detail than that available in either imaging modality operating independently.
There are two major advantages of implementing this new technology to guide surgeries and interventional procedures. First, the accurate target localization and acceptable target visibility reported in Table 2 and Sections 3.2 and 3.3 indicate that the flexible array provides complete contact without introducing extra applied pressure and sacrificing target visibility. Therefore, frequent pressure on the organ or tissue surfaces can be avoided to reduce target displacement, patient discomfort, and possible tissue damage. The ex vivo results reported in Section 3.4 further demonstrate the promise of using the flexible array for photoacoustic-guided surgeries or interventions. Second, it is advantageous that the element position calculation and time delay computation derived for this study are applicable not only for photoacoustic imaging but also for ultrasound imaging. The flexible array transducer used in this work was designed to be capable of both transmitting and receiving signals. Therefore, the ultrasound and photoacoustic raw data can be acquired simultaneously. The ultrasound images shown in Figs. 7 and 12 were reconstructed using Eqs. (1)–(6) for both transmitting and receiving processes. The co-registered ultrasound image provides qualitative information about the surrounding tissue when inserting the surgical needle during the interventional operation, as demonstrated in Fig. 12, with the laparoscopic images offering additional confirmation of structural content.
Our results depend on a priori knowledge of the flexible array curvature, which can possibly be obtained for surgical guidance using preoperative images or tools like radius gauges [45], followed by calculations of element positions using Eqs. (1)–(5). To maintain complete contact with the calculated curved surface for the duration of a procedure, the array can be taped to the surface. As an alternative, real-time array element position monitoring may be necessary in more dynamic environments (e.g., regular organ motion and deformation, influences from cardiac or respiratory cycles). In these cases, techniques such as optical trackers [26,34] or shape estimation algorithms [24,25,27] can be applied to track element positions and estimate transducer geometry in more complicated shapes.
While we successfully reconstructed photoacoustic images with the fiber tip at the correct positions, we also observed target size deviations relative to ground truth values and target visibility variations. When comparing the target sizes measured in the third and fourth rows in Table 2 and in Figs. 11(a) and 13(a), the observed deviations have two possible causes, outside of potential shape estimation errors. First, the optical fiber translation might not have been perfectly aligned with the elevational axis of the transducer. Second, the signal might incorporate the metal needle surrounding the optical fiber. The needle used in this study has a diameter of 2 mm, which is larger than those of the core (600 $\mu$m) and coating (1040 $\mu$m) of the optical fiber. The observed target visibility variations and outliers for each stationary optical fiber position are likely to be caused by expected energy fluctuations throughout the experiment [46,47]. These variations and outliers can possibly be reduced with real-time energy compensation if greater stability is required for an interventional procedure.
Future work can challenge this new flexible array technology by incorporating irregular transducer geometries. The impact of approximating an irregular surface geometry as a radial surface is demonstrated in Figs. 5–6. Improvements may be achieved with additional innovations in shape estimation beamformers and image reconstruction methods. In addition, it may also be beneficial to modify image reconstruction methods for less complicated surface geometries, as the conventional DAS beamformer employed in this work produced noticeable artifacts in the presence of tissue (see Fig. 12). Potential options to remove these artifacts include short-lag spatial coherence beamforming [39,48] and deep learning [49].
5. Conclusion
This work is the first to present multiple demonstrations of photoacoustic imaging with a new flexible array transducer for potential surgical or interventional applications. We establish the feasibility of deploying our new flexible array, targeting cases that require visualization of interventional tool tips (e.g., needle or catheter tips augmented with optical fibers). Results are particularly applicable when the array geometry is known through sources such as preoperative images (e.g., CT or MRI of the abdominal surface) or through optical trackers. Equations are included to estimate element positions from known geometry, and these equations are demonstrated to be applicable for both ultrasound and photoacoustic images acquired with the same flexible array transducer. Feasibility is demonstrated and validated with a simulation study, image reconstruction error analysis, and five phantoms of known geometry (including one phantom containing ex vivo liver tissue to present anatomical expectations in addition to photoacoustic images of an interventional tool tip). Standard image quality assessments for target visibility, lateral and axial target sizing, and lateral and axial target position accuracy are presented. These quantitative results provide a complete initial characterization of future potential, and they are within acceptable ranges for photoacoustic imaging technology.
Funding
National Science Foundation (1751522, 2014088).
Acknowledgments
This work was supported by NSF CAREER Award ECCS-1751522 and NSF SCH Award IIS-2014088. The authors thank Mardava Gubbi for discussions about the conversion of coordinate systems.
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. M. A. Lediju Bell, “Photoacoustic imaging for surgical guidance: Principles, applications, and outlook,” J. Appl. Phys. 128(6), 060904 (2020). [CrossRef]
2. A. Wiacek and M. A. L. Bell, “Photoacoustic-guided surgery from head to toe,” Biomed. Opt. Express 12(4), 2079–2117 (2021). [CrossRef]
3. T. Zhao, A. E. Desjardins, S. Ourselin, T. Vercauteren, and W. Xia, “Minimally invasive photoacoustic imaging: Current status and future perspectives,” Photoacoustics 16, 100146 (2019). [CrossRef]
4. M. Xu and L. V. Wang, “Photoacoustic imaging in biomedicine,” Rev. Sci. Instrum. 77(4), 041101 (2006). [CrossRef]
5. P. Beard, “Biomedical photoacoustic imaging,” Interface Focus 1(4), 602–631 (2011). [CrossRef]
6. L. V. Wang, “Tutorial on photoacoustic microscopy and computed tomography,” IEEE J. Sel. Top. Quantum Electron. 14(1), 171–179 (2008). [CrossRef]
7. L. Giovanella, L. Bacigalupo, G. Treglia, and A. Piccardo, “Will 18 f-fluorocholine pet/ct replace other methods of preoperative parathyroid imaging?” Endocrine 71(2), 285–297 (2021). [CrossRef]
8. H. W. Kim, J.-C. Lee, K.-H. Paik, J. Kang, Y. H. Kim, Y.-S. Yoon, H.-S. Han, J. Kim, and J.-H. Hwang, “Adjunctive role of preoperative liver magnetic resonance imaging for potentially resectable pancreatic cancer,” Surgery 161(6), 1579–1587 (2017). [CrossRef]
9. D. Piras, C. Grijsen, P. Schütte, W. Steenbergen, and S. Manohar, “Photoacoustic needle: minimally invasive guidance to biopsy,” J. Biomed. Opt. 18(7), 070502 (2013). [CrossRef]
10. J. Shubert and M. A. L. Bell, “Photoacoustic based visual servoing of needle tips to improve biopsy on obese patients,” in 2017 IEEE International Ultrasonics Symposium (IUS), (IEEE, 2017), pp. 1–4.
11. M. A. Lediju Bell and J. Shubert, “Photoacoustic-based visual servoing of a needle tip,” Sci. Rep. 8(1), 15519 (2018). [CrossRef]
12. M. Graham, F. Assis, D. Allman, A. Wiacek, E. Gonzalez, M. Gubbi, J. Dong, H. Hou, S. Beck, J. Chrispin, and M. A. L. Bell, “In vivo demonstration of photoacoustic image guidance and robotic visual servoing for cardiac catheter-based interventions,” IEEE Trans. Med. Imaging 39(4), 1015–1029 (2020). [CrossRef]
13. B. Eddins and M. A. L. Bell, “Design of a multifiber light delivery system for photoacoustic-guided surgery,” J. Biomed. Opt. 22(04), 1 (2017). [CrossRef]
14. J. Shubert and M. A. L. Bell, “A novel drill design for photoacoustic guided surgeries,” in Photons Plus Ultrasound: Imaging and Sensing 2018, vol. 10494 (SPIE, 2018), pp. 38–43.
15. J. E. Aldrich, “Basic physics of ultrasound imaging,” Crit. Care Med. 35(Suppl), S131–S137 (2007). [CrossRef]
16. M. Blaivas, M. Lyon, L. Brannam, S. Duggal, and P. Sierzenski, “Water bath evaluation technique for emergency ultrasound of painful superficial structures,” The Am. J. Emerg. Med. 22(7), 589–593 (2004). [CrossRef]
17. K. Nakahata, S. Tokumasu, A. Sakai, Y. Iwata, K. Ohira, and Y. Ogura, “Ultrasonic imaging using signal post-processing for a flexible array transducer,” NDT & E Int. 82, 13–25 (2016). [CrossRef]
18. M. T. Graham, J. Huang, F. X. Creighton, and M. A. L. Bell, “Simulations and human cadaver head studies to identify optimal acoustic receiver locations for minimally invasive photoacoustic-guided neurosurgery,” Photoacoustics 19, 100183 (2020). [CrossRef]
19. J. Shubert and M. A. L. Bell, “Photoacoustic imaging of a human vertebra: implications for guiding spinal fusion surgeries,” Phys. Med. Biol. 63(14), 144001 (2018). [CrossRef]
20. E. A. Gonzalez, A. Jain, and M. A. L. Bell, “Combined ultrasound and photoacoustic image guidance of spinal pedicle cannulation demonstrated with intact ex vivo specimens,” IEEE Trans. Biomed. Eng. 68(8), 2479–2489 (2021). [CrossRef]
21. K. M. Kempski, A. Wiacek, M. Graham, E. González, B. Goodson, D. Allman, J. E. Palmer, H. Hou, S. Beck, J. He, and M. A. L. Bell, “In vivo photoacoustic imaging of major blood vessels in the pancreas and liver during surgery,” J. Biomed. Opt. 24(12), 1 (2019). [CrossRef]
22. A. Wiacek, K. C. Wang, H. Wu, and M. A. L. Bell, “Photoacoustic-guided laparoscopic and open hysterectomy procedures demonstrated with human cadavers,” IEEE Trans. Med. Imaging 40(12), 3279–3292 (2021). [CrossRef]
23. C. J. Lane, “The inspection of curved components using flexible ultrasonic arrays and shape sensing fibres,” Case Stud. Nondestruct. Test. Eval. 1, 13–18 (2014). [CrossRef]
24. J. Chang, Z. Chen, Y. Huang, Y. Li, X. Zeng, and C. Lu, “Flexible ultrasonic array for breast-cancer diagnosis based on a self-shape–estimation algorithm,” Ultrasonics 108, 106199 (2020). [CrossRef]
25. T. Noda, N. Tomii, K. Nakagawa, T. Azuma, and I. Sakuma, “Shape estimation algorithm for ultrasound imaging by flexible array transducer,” IEEE Trans. Ultrason., Ferroelect., Freq. Contr. 67(11), 2345–2353 (2020). [CrossRef]
26. D. China, Z. Feng, H. Hooshangnejad, D. Sforza, P. Vagdargi, M. A. L. Bell, A. Uneri, and K. Ding, “Real-time element position tracking of flexible array transducer for ultrasound beamforming,” in Ultrasonic Imaging and Tomography, (SPIE, 2023).
27. T. Noda, T. Azuma, Y. Ohtake, I. Sakuma, and N. Tomii, “Ultrasound imaging with a flexible probe based on element array geometry estimation using deep neural network,” IEEE Trans. Ultrason., Ferroelect., Freq. Contr. 69(12), 3232–3242 (2022). [CrossRef]
28. X. Huang, M. A. L. Bell, and K. Ding, “Deep learning for ultrasound beamforming in flexible array transducer,” IEEE Trans. Med. Imaging 40(11), 3178–3189 (2021). [CrossRef]
29. Q. T. K. Shea, Y. T. Ling, T. T.-Y. Lee, and Y. P. Zheng, “Spinal deformity measurement using a low-density flexible array ultrasound transducer: A feasibility study with phantoms,” Med. Nov. Technol. Devices 11, 100090 (2021). [CrossRef]
30. K. Roy, S. Agrawal, A. Dangi, T. Liu, H. Chen, T. N. Jackson, R. Pratap, and S.-R. Kothapalli, “Body conformal linear ultrasound array for combined ultrasound and photoacoustic imaging,” in 2020 IEEE International Ultrasonics Symposium (IUS), (IEEE, 2020), pp. 1–4.
31. M. Ghavami, A. K. Ilkhechi, and R. Zemp, “Flexible transparent CMUT arrays for photoacoustic tomography,” Opt. Express 30(10), 15877–15894 (2022). [CrossRef]
32. J. Zhang, A. Wiacek, E. González, Z. Feng, K. Ding, and M. A. L. Bell, “A flexible array transducer for photoacoustic-guided surgery,” in 2022 IEEE International Ultrasonics Symposium (IUS), (IEEE, 2022), pp. 1–4.
33. J. Zhang, A. Wiacek, Z. Feng, K. Ding, and M. A. L. Bell, “Comparison of flexible array with laparoscopic transducer for photoacoustic-guided surgery,” in Photons Plus Ultrasound: Imaging and Sensing 2023, vol. 12379 (SPIE, 2023), pp. 205–212.
34. X. Huang, H. Hooshangnejad, D. China, Z. Feng, J. Lee, M. A. L. Bell, and K. Ding, “Ultrasound imaging with flexible array transducer for pancreatic cancer radiation therapy,” Cancers 15(13), 3294 (2023). [CrossRef]
35. X. Li, “Ultrasound scan conversion on ti’s c64x+ dsps,” Application Report SPRAB32, Texas Instruments (2009).
36. B. E. Treeby and B. T. Cox, “k-wave: Matlab toolbox for the simulation and reconstruction of photoacoustic wave fields,” J. Biomed. Opt. 15(2), 021314 (2010). [CrossRef]
37. W. C. Vogt, C. Jia, K. A. Wear, B. S. Garra, and T. Joshua Pfefer, “Biologically relevant photoacoustic imaging phantoms with tunable optical and acoustic properties,” J. Biomed. Opt. 21(10), 101405 (2016). [CrossRef]
38. G. M. Spirou, A. A. Oraevsky, I. A. Vitkin, and W. M. Whelan, “Optical and acoustic properties at 1064 nm of polyvinyl chloride-plastisol for use as a tissue phantom in biomedical optoacoustics,” Phys. Med. Biol. 50(14), N141–N153 (2005). [CrossRef]
39. M. A. L. Bell, N. Kuo, D. Y. Song, and E. M. Boctor, “Short-lag spatial coherence beamforming of photoacoustic images for enhanced visualization of prostate brachytherapy seeds,” Biomed. Opt. Express 4(10), 1964–1977 (2013). [CrossRef]
40. E. Jackson, C. Coussios, and R. Cleveland, “Nonlinear acoustic properties of ex vivo bovine liver and the effects of temperature and denaturation,” Phys. Med. Biol. 59(12), 3223–3238 (2014). [CrossRef]
41. K. M. Kempski, M. T. Graham, M. R. Gubbi, T. Palmer, and M. A. L. Bell, “Application of the generalized contrast-to-noise ratio to assess photoacoustic image quality,” Biomed. Opt. Express 11(7), 3684–3698 (2020). [CrossRef]
42. A. Rodriguez-Molares, O. M. H. Rindal, J. D’hooge, S.-E. Måsøy, A. Austeng, M. A. L. Bell, and H. Torp, “The generalized contrast-to-noise ratio: A formal definition for lesion detectability,” IEEE Trans. Ultrason., Ferroelect., Freq. Contr. 67(4), 745–759 (2020). [CrossRef]
43. M. R. Gubbi, E. A. Gonzalez, and M. A. L. Bell, “Theoretical framework to predict generalized contrast-to-noise ratios of photoacoustic images with applications to computer vision,” IEEE Trans. Ultrason., Ferroelect., Freq. Contr. 69(6), 2098–2114 (2022). [CrossRef]
44. M. A. L. Bell, A. K. Ostrowski, K. Li, P. Kazanzides, and E. M. Boctor, “Localization of transcranial targets for photoacoustic-guided endonasal surgeries,” Photoacoustics 3(2), 78–87 (2015). [CrossRef]
45. H. Kurosawa, P. Walker, S. Abe, A. Garg, and T. Hunter, “Geometry and motion of the knee for implant and orthotic design,” J. Biomech. 18(7), 487–499 (1985). [CrossRef]
46. A. Bartana, R. Kosloff, and D. J. Tannor, “Laser cooling of molecular internal degrees of freedom by a series of shaped pulses,” The J. Chem. Phys. 99(1), 196–210 (1993). [CrossRef]
47. G. Wang, E. M. Sevick, E. Mittag, D. J. Searles, and D. J. Evans, “Experimental demonstration of violations of the second law of thermodynamics for small systems and short time scales,” Phys. Rev. Lett. 89(5), 050601 (2002). [CrossRef]
48. M. T. Graham and M. A. L. Bell, “Photoacoustic spatial coherence theory and applications to coherence-based image contrast and resolution,” IEEE Trans. Ultrason., Ferroelect., Freq. Contr. 67(10), 2069–2084 (2020). [CrossRef]
49. D. Allman, A. Reiter, and M. A. L. Bell, “Photoacoustic source detection and reflection artifact removal enabled by deep learning,” IEEE Trans. Med. Imaging 37(6), 1464–1477 (2018). [CrossRef]
