1214 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 43, NO.

3, MARCH 2024

Moving Beyond Simulation: Data-Driven

Quantitative Photoacoustic Imaging
Using Tissue-Mimicking Phantoms
Janek Gröhl , Thomas R. Else , Lina Hacker , Ellie V. Bunce , Paul W. Sweeney,
and Sarah E. Bohndiek

Abstract — Accurate measurement of optical absorption depth-dependent loss of signal intensity. We demonstrate
coefficients from photoacoustic imaging (PAI) data would that training on experimental phantoms can restore the
enable direct mapping of molecular concentrations, pro- correlation of signal amplitudes measured in depth. While
viding vital clinical insight. The ill-posed nature of the the absolute quantification error remains high and further
problem of absorption coefficient recovery has prohibited improvements are needed, our results highlight the promise
PAI from achieving this goal in living systems due to the of deep learning to advance quantitative PAI.
domain gap between simulation and experiment. To bridge
this gap, we introduce a collection of experimentally well- Index Terms— Deep learning, image reconstruction, pho-
characterised imaging phantoms and their digital twins. toacoustics, quantitative imaging.
This first-of-a-kind phantom data set enables supervised
I. I NTRODUCTION
training of a U-Net on experimental data for pixel-wise
estimation of absorption coefficients. We show that train-
ing on simulated data results in artefacts and biases in
the estimates, reinforcing the existence of a domain gap
P HOTOACOUSTIC (PA) imaging (PAI) has the unique
ability to provide high-resolution optical-contrast molec-
ular imaging up to several centimetres deep in tissue with
between simulation and experiment. Training on experimen- potential clinical application areas ranging from diagnosis of
tally acquired data, however, yielded more accurate and
robust estimates of optical absorption coefficients. We com- breast cancer [1] or cardiovascular disease [2], to grading of
pare the results to fluence correction with a Monte Carlo skin [3], [4] or inflammatory diseases [5]. PAI transforms
model from reference optical properties of the materials, absorbed laser energy into acoustic sound waves [6], [7]
which yields a quantification error of approximately 20%. that can be measured as time-series pressure data p(t) using
Application of the trained U-Nets to a blood flow phantom acoustic detectors [8], [9].
demonstrated spectral biases when training on simulated
data, while application to a mouse model highlighted the A key aim of PAI is estimation of absolute molecular con-
ability of both learning-based approaches to recover the centrations from images acquired at multiple wavelengths [10],
[11]. To achieve this goal, two inverse problems have to
Manuscript received 18 September 2023; accepted 5 November 2023. be solved: the acoustic inverse problem, which refers to the
Date of publication 8 November 2023; date of current version 4 March
2024. This work was supported by the NVIDIA Academic Hardware Grant reconstruction of the initial pressure distribution p0 from
Program and utilized two Quadro Ray Tracing Texel eXtreme (RTX) 8000. p(t) and the optical inverse problem, which refers to the
The work of Janek Gröhl was supported by the Deutsche Forschungs- estimation of the optical absorption coefficient µa from p0 .
gemeinschaft (DFG, German Research Foundation) under project GR
5824/1. The work of Lina Hacker was supported by the MedAccel Pro- Reconstructing p0 from p(t) requires an inverse acoustic
gramme of National Physical Laboratory (NPL) financed by the Industrial operator, denoted as A→1 (an image reconstruction algorithm).
Strategy Challenge Fund of the Department for Business, Energy and To obtain accurate reconstructions, the acoustic properties of
Industrial Strategy. The work of Thomas R. Else, Paul W. Sweeney,
and Sarah E. Bohndiek was supported by the Cancer Research U.K. the tissue, ω, and the characteristics of the acoustic detector,
under Grant C9545/A29580. The work of Ellie V. Bunce was supported ε, must be taken into consideration: p0 ↑ A→1 ( p(t, ε), ω, ε).
by the Cancer Research U.K. (Radiation Research Network (Rad- Once reconstructed, determining µa from p0 , for a given
Net) Cambridge) under Grant C17918/A28870. (Corresponding author:
Janek Gröhl.) wavelength ω at position x, must deal with the fact that
This work involved human subjects or animals in its research. Approval p0 depends not only on the local absorption coefficient µa ,
of all ethical and experimental procedures and protocols was granted by but also the temperature (T )-dependent Grüneisen parameter
the Project and Personal Licences under PPL No. PE12C2B96, issued
under the United Kingdom Animals (Scientific Procedures) Act, 1986 and ϑ and the light fluence ϖ, which is a function of both µa and
approved locally under compliance form CFSB2317. the optical scattering coefficient µs : p0 (x, ω, T ) = ϑ(x, T ) ·
Janek Gröhl, Thomas R. Else, Ellie V. Bunce, Paul W. Sweeney, and µa (x, ω) · ϖ(x, ω, µa (ω), µs (ω)).
Sarah E. Bohndiek are with the Cancer Research U.K. Cambridge
Institute, CB2 0RE Cambridge, U.K., and also with the Department The spatial distribution of ϖ and ϑ in living subjects are
of Physics, University of Cambridge, CB3 0HE Cambridge, U.K. typically unknown but must be accounted for to accurately
(e-mail: [email protected]; [email protected]; [email protected]; determine µa . ϑ is often assumed constant [11]. ϖ could be
[email protected]; [email protected]).
Lina Hacker was with the Department of Physics and the accounted for by iterative methods, using fixed-point iteration
Cancer Research U.K. Cambridge Institute, University of Cambridge, schemes assuming a priori known µs [12] or variational
CB3 0HE Cambridge, U.K. She is now with the Department of approaches, allowing a range of constraints and priors [13].
Oncology, University of Oxford, OX3 7DQ Oxford, U.K. (e-mail:
[email protected]). To date, these methods have been restricted to simulated data,
Digital Object Identifier 10.1109/TMI.2023.3331198 except when including a calibrated imaging target into the

© 2023 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/
GRÖHL et al.: MOVING BEYOND SIMULATION: DATA-DRIVEN QUANTITATIVE PHOTOACOUSTIC IMAGING 1215

medium [14]. Deep learning approaches have been proposed and µa = 0.5 cm→1 – 4.0 cm→1 in the inclusions. Reduced
to solve both the acoustic [15], [16] and optical inverse prob- scattering was set to be µ↔s = 5 cm→1 – 15 cm→1 , defined at
lems [17], [18], [19] in PAI. Iterative or variational methods a reference wavelength of 800 nm. We attempted to achieve
typically work well with simulated data, but do not achieve a uniform distribution over this range for the inclusions, but
convincing results on experimental data [20]. A key limitation due to experimental constraints and the exclusion of phantoms
for application in living subjects is that the optical and acoustic from a manual quality assurance step, this was not always
properties are unknown, making validation of quantitative PAI possible. For the background we initially created a uniform
methods extremely difficult. One indirect possibility to validate distribution of absorption coefficients and complemented this
these methods is to analyse the performance of methods that later on with more phantoms with a background absorption
use the estimated µa as an input, for example, semantic image of 0.1 cm→1 when extending the dataset, leading to a skewed
segmentation [21], [22], or oximetry [23]. Li et al. [24] have distribution. The phantoms featured various inclusion patterns
recently proposed an approach to enable training on simulated with distinct µa and µ↔s combinations in both the inclusions
data with known ground truth, using a cycle-GAN network to and the backgrounds.
transform the simulation data distribution into the experimental
domain, enabling them to train a network on transformed B. Optical Property Characterisation
simulated data and apply it to previously unseen experimental For each phantom material, rectangular samples with a
data. length of 5.9 cm, a width of 1.8 cm, and a thickness ranging
Here, we bridge the gap from simulated to experimental between 2 mm and 4 mm were prepared for optical character-
training data by fabricating a collection of well-characterised isation. Sample thicknesses were determined at five distinct
co-polymer-in-oil tissue-mimicking phantoms [25] covering locations using digital vernier calipers. We used an in-house
a biologically relevant range of µa and µs , to enable the use double integrating sphere (DIS) system [25] based on the
of supervised deep learning for the optical inverse problem system developed by Pickering et al. [28] to determine µa and
of PAI. We create digital phantom twins by characterising µ↔s of the phantom material in a wavelength range of 600 to
optical properties with a double integrating sphere (DIS) 950 nm. The recorded transmittance and reflectance values
system and train U-Nets [26] to estimate µa on matching were normalised and analysed with the inverse adding dou-
simulated and experimental PA images. We apply the trained bling (IAD) algorithm [29] to estimate the optical properties
U-Nets to a held-out set of test phantoms and compare the of the material. The anisotropy (g) and refractive index (n)
results to fluence correction with a Monte Carlo model based were assumed to be g = 0.7 and n = 1.4 [30].
on the DIS reference measurements. Importantly, we then
explore the potential of the methods on pre-clinical data from C. Photoacoustic Imaging and Data Processing
living subjects. The phantoms were imaged using a pre-clinical PAI sys-
tem (MSOT inVision-256TF, iThera Medical GmbH, Munich,
II. M ETHODS Germany). It has 256 transducer elements in a circular array
A. Tissue Mimicking Phantoms with an angular coverage of 270↓ which have a 5-MHz
centre frequency and 60% bandwidth. More details on the
A total of 137 cylindrical phantoms with a diameter of
measurement setup can be found in a prior publication [31].
27.5 mm, a height of 65-80 mm, and an approximate volume
We imaged in the wavelength range from 700 nm to 900 nm
of 40-50 mL were fabricated based on a previously published
in 10 nm steps using 10 frame averaging. The phantoms were
protocol [25]. Briefly, 83.80 g mineral oil (Sigma Aldrich-
mounted in the imaging system using a custom 3D-printed
330779-1L) was used as a base. 25.14 g polystyrene-block-
polylactic acid phantom holder printed with the Prusa i3
poly(ethylene-ran-butylene)-block-polystyrene (SEBS, Sigma
MK3S+ (Prusa Research, Czechia) and immersed in deionized
Aldrich 200557-250G) as the polymer and 2.09 g butylated
water for acoustic coupling. The phantoms were imaged at
hydroxytoluene (HT, Sigma Aldrich W218405-1KG-K) as an
different cross-sections in 2 mm steps along a 44 mm-long
antioxidant was added to the mineral oil base, which was
section. The temperature of the water bath was set to 25 ↓ C.
heated to 150 ↓ C and poured into a mould (for details see [25]
Images were reconstructed using the delay-and-sum
and [27]). Different concentrations of titanium dioxide (TiO2 ,
algorithm with the PATATO toolbox [32] with a bandpass filter
Sigma Aldrich 232033-100g) and alcohol-soluble nigrosin
between 5 KHz and 7 MHz, application of a Hilbert transform,
(Sigma Aldrich 211680-100G) were used to tune µ↔s and µa
as well as time interpolation by a factor of three and detection
respectively. Each phantom consisted of a background cylinder
element interpolation by a factor of two. The images were
into which imaging targets (referred to as inclusions) were
reconstructed into a 300 ↗ 300 pixel grid with a field of
added. The background shape was created using the cut body
view of 32 ↗ 32 mm. To enable compatibility with the U-Net,
from a glass syringe (Sigma Aldrich CADG5157). Addition
images were cropped to 288 ↗ 288 pixels without changing
of the inclusions was performed using either careful pouring
the resolution (Fig. 1 A).
(diameter > 6 mm) or via a vacuum pump. Space for the inclu-
sions was created by either using metal insertion rods with
varying diameters or by using microscopy glass slides to shield D. Data Quality Assurance
an area of the glass mould. Inclusions were prepared in such A wide range of artefacts can corrupt the phantom prepa-
a way as to be z-invariant. Optical properties were sampled ration and imaging process when creating a wide range of
from the ranges µa = 0.05 cm→1 – 0.5 cm→1 in the background inclusion geometries. Sources of artefacts include: unsolved
1216 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 43, NO. 3, MARCH 2024

(last accessed 04/04/2023). The detection geometry consists of

256 toroidally focussed transducer elements covering a 270 ↓
field of view around the imaging target. The arrangement of
the transducers was implemented in the k-Wave toolbox [35].
Each segmentation class was assigned the optical properties
determined by the DIS measurements for the sample material
at 21 wavelengths from 700 nm to 900 nm in 10 nm steps. The
background medium was assumed to be pure water [36], as this
is the acoustic couplant used in the imaging system. We sim-
ulated the forward models using the SIMPA toolkit [37].
The light distribution was simulated in 3D using the MCX
Monte Carlo model and the propagation of sound waves was
simulated in 2D using a custom k-Wave script (Fig. 1 B). For
computational efficiency, we limited the Monte Carlo model
to 107 photons for each simulation run and only simulated the
acoustic model in 2D assuming an infinitely symmetric initial
pressure outside of the imaging plane. Despite all assumptions,
the simulations correlate well with the experiment (Pearson
correlation coefficient R=0.94).

F. Training and Held-Out Test Data

The 102 phantoms that passed the quality assurance step
were separated into a training set with 84 phantoms and a test
set with 18 phantoms, roughly corresponding to the typically
used 80/20 data split in deep learning. Since we imaged at
21 distinct wavelengths, we had 1764 training and 378 test
data images. Test data was derived from a held out set of
phantoms which did not contain base material µa and µ↔s value
pairs that were used in the training set. Furthermore, ten of
Fig. 1. Supervised deep learning approach. A library of tissue-
mimicking phantoms with known optical properties were subjected to:
the test phantoms contained inclusion geometries not present
(A) Experimental photoacoustic imaging (PAI), followed by delay-and- in the training data.
sum reconstruction, manual image quality assessment, and segmenta-
tion; and (B) Computation of light fluence and time series pressure using
digital twins of the phantom and imaging system. (C) Paired experimental G. Absorption Estimation Methods
and computational data were used to train a supervised deep learning
model to estimate optical absorption coefficients.
We applied four methods for calibration of the optical
absorption coefficient µa against the PAI data (Fig. 2).
Linear calibration was chosen as a baseline model. Lin-
polymers; material inhomogeneities; air bubbles; yellowing ear calibration is highly inaccurate, as it completely ignores
from oxidation; and acoustic reflections from the 3D-printed the dependency of the PA signal on light fluence or arte-
phantom holder. After image reconstruction, manual quality facts introduced by image reconstruction. Unfortunately, this
control was performed for each tomographic cross-section to method is commonly used in literature when applying linear
check for the presence of artefacts that impacted data quality. unmixing techniques for sO2 estimation directly to the recon-
Phantoms were only included in subsequent analysis if at least structed images. To circumvent depth-dependent attenuation,
one cross-sectional slice contained no artefacts. Using this we selected the brightest 2% of pixels in the background
process, 35 of the 137 phantoms were excluded. material of the training phantoms (i.e. the rim) and applied
a linear regression to relate the arbitrary units PA signals in
these pixels to the reference optical absorption (Cal., Fig. 2A),
E. Synthetic Imaging With Digital Twins giving µa = 14851
( pˆ0 → 313), R=0.94.
We manually segmented the cross-sectional images using Fluence correction with reference Monte Carlo simu-
the Medical Imaging Interaction Toolkit (MITK) [33] into lations was performed as a gold standard method. Monte
segmentation classes of background and target inclusions, Carlo simulations of light fluence were made based on the
which were used for creating the digital phantom twins. reference absorption and scattering values obtained by the DIS
We implemented a digital twin of the PAI device including measurements (GT-ϖ, Fig. 2B). The resulting linear fit on the
both the illumination and detection geometry. The light source fluence corrected data gave µa = 8801 1
( ϖpMC
ˆ0
→ 832), R=0.93.
consists of five fibre bundle pairs radially distributed around Absorption estimation with deep learning was achieved
the object and was implemented as a custom illumination in by training a modified U-Net [26] in a supervised manner on
Monte Carlo eXtreme (MCX) [34]. The implementation of the training data (Fig. 1 B). The U-Net was used without a final
this system can be found in https://github.com/jgroehl/mcx activation layer, to make it suitable for regression tasks [18].
GRÖHL et al.: MOVING BEYOND SIMULATION: DATA-DRIVEN QUANTITATIVE PHOTOACOUSTIC IMAGING 1217

We trained two U-Nets: one on the simulated dataset

(DL-Sim, Fig. 2C) and one on the experimentally acquired
training data set (DL-Exp, Fig. 2D). The training data
were normalised by the mean and standard deviation of the
respective training data set. The networks were trained for
200 epochs with an initial learning rate of 10→4 and a learning
rate scheduler that reduced the learning rate by 10% on a
plateau of the validation loss for more than 10 epochs. We used
an ensemble of five U-Nets trained with five-fold cross-
validation where each fold had 16 distinct training phantoms
as a validation set.
After training, absorption estimates µˆa were made on
unseen experimental PA images pˆ0 , after normalising by
the mean and standard deviations of the entire ! experimental
"
pˆ →µ
training data set (µExp and ϱExp ): µˆa = DL 0ϱExpExp .

H. Performance Evaluation
For the final performance evaluation, the estimates of each
of the five training folds were averaged. In addition to per-
formance evaluation on the held-out test phantoms, we also
applied the network to completely different experimental data
without additional training. To this end, we used a blood
flow phantom experiment and in vivo imaging of healthy
mice. While these experiments allow us to investigate the
generalisability and limitations of the models, inaccuracies
are expected due to variability in the Grüneisen parameter
distribution compared to the phantoms.
1) Flow Phantom Data: A variable oxygenation blood flow
phantom was imaged as described previously [38]. We fabri-
cated agar-based cylindrical phantoms with a radius of 10 mm
and embedded a polyvinyl chloride tube (i.d. 0.8 mm, o.d.
0.9 mm) in the centre of the phantom during the agar setting
process at room temperature. For the base mixture, 1.5 % (w/v)
agar (Sigma Aldrich 9012-36-6) was added to Milli-Q water
and was heated until it dissolved. Once the mixture had cooled
to approximately 40↓ C, 2.08% (v/v) of pre-warmed intralipid
(20% emulsion, Merck, 68890-65-3) and 0.74% (v/v) Nigrosin
solution (0.5 mg/ml in Milli-Q water) was added, mixed, and
Fig. 2. Optical absorption estimation methods. (A) The baseline poured into the mould. PA images of the phantom were taken
model (Cal.) is a linear mapping from the arbitrary units PA image to between 700 nm and 900 nm in 20 nm steps.
reference absorption units. (B) Adding complexity, the simulated light
fluence is used to correct PA images (GT-!) before a linear mapping is
2) In Vivo Mouse Data: All animal procedures were con-
applied. (C) For supervised deep learning, a U-Net is trained on simulated ducted in accordance with project and personal licences (PPL
input data to estimate the optical absorption based on the input PAI data number: PE12C2B96), issued under the United Kingdom Ani-
(DL-Sim) and (D) trained on experimentally acquired data (DL-Exp).
mals (Scientific Procedures) Act, 1986 and approved locally
under compliance form CFSB2317. Nine healthy 9-week-
old female C57BL/6 albino mice were imaged according to
a previously established standard operating procedure [31].
To make use of the digital twin simulations during training, Imaging was performed at 10 wavelengths between 700 and
we defined the U-Net to have a two-channel output: the first 900 nm averaging over 10 scans each. The kidneys, spleen,
channel to estimate µa and the second channel to estimate ϖ. spine, and aorta were segmented in PA images using MITK.
Our U-Net implementation used strided convolutions with a 3) Performance Measures: In this work, we report the rela-
(3 ↗ 3) kernel size for downsampling and strided transpose tive error, which is defined as Rel. Error(x, x̂) = |x→x x̂| ↘ 100,
convolutions with a (2 ↗ 2) kernel size for upsampling. It used and the absolute error, which is defined as Abs. Error(x, x̂) =
leaky rectified linear units as the non-linear activation layers |x → x̂|, where x̂ is the estimate for a quantity of interest
and has four sampling steps with a total of 1024 channels in x. We compute the mean of the pixel-wise measure for
the bottleneck and used the mean squared error as the loss each segmentation region. When reporting aggregate results,
function for both output channels. we proceed to compute the median and interquartile range
1218 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 43, NO. 3, MARCH 2024

Fig. 3. Representative performance examples from the held-out test set. (A, F, K) Baseline model (Cal., yellow); (B, G, L) Fluence correction
with reference simulations (GT-ω, green); (C, H, M) U-Net trained on simulated data (DL-Sim, blue); and (D, I, N) U-Net trained on experimental data
(DL-Exp, red). (E, J, O) show line profiles with the respective colours. The methods are colour coded in their dedicated colours and the ground truth
optical absorption is shown in a solid black line (GT µa ). The level windows are adjusted to clearly visualise the water background, the background
material, and the inclusions.

of the resulting distribution of mean errors per segmentation J. Statistics

region. We also report the generalised # Ncontrast-to-noise ratio, The degree of linear correlation between variables was
→1
which is defined as gCNR = 1 → k=0 min(h i (k), h b (k)), analysed using the Pearson correlation coefficient. Statistical
where h i and h b are the histograms of the signals in the significance is tested using a non-parametric unpaired Mann-
inclusion and background and k is the index of the histogram Whitney U test. The significance is indicated by * (P < 0.05),
bin [39]. The average in a certain background segmentation ** (P < 0.01), *** (P < 0.001) and **** (P < 0.0001). Data
class was taken as the mean, while the average in a inclusion are reported as mean ± standard deviation unless otherwise
class was the median over the first 1.28 mm depth inside stated.
the inclusions. The depth threshold was manually determined
sweeping from 0 mm to 5 mm depth to yield the best µa
quantification results on the training data when using the GT-ϖ III. R ESULTS
model. GT-ϖ is the model expected to be most affected by the A. Deep Learning Models Show Improved Spatial
depth-dependent signal decrease. Distribution of Absorption Estimates
We first inspected the images of estimated µa over all test
I. Oxygenation Estimation phantoms. We highlight here three examples representative of
For the phantom and in vivo data analysis the blood oxy- different observations at 800 nm, illustrated using a line profile
genation (s O2 ) was calculated using linear spectral unmixing through the inclusions (Fig. 3). The first example shows that
(LU). We used oxy- and deoxyhemoglobin as our endmem- GT-ϖ can overestimate µa deep inside the test object, while
bers with reference spectra from literature [40] and solved DL-Exp can reconstruct a smooth and homogeneous back-
the system of linear equations with the pseudo inverse as ground (Fig. 3A-E). The second example shows a challenging
implemented in the PATATO toolkit [32]. test case (Fig. 3F-J) with a halo artefact around one inclusion
GRÖHL et al.: MOVING BEYOND SIMULATION: DATA-DRIVEN QUANTITATIVE PHOTOACOUSTIC IMAGING 1219

where DL-Exp achieved a good agreement with the reference

in the background (DL-Exp: 0.12 ± 0.07 cm→1 vs Reference:
0.098 cm→1 ) and one inclusion (DL-Exp: 1.0 ± 0.12 cm→1
vs Reference: 1.04 cm→1 ) but showed poorer performance in
the other inclusion (DL-Exp: 1.97 ± 0.35 cm→1 vs Reference:
3.1 cm→1 ). Finally, we show an out-of-distribution phantom
with a geometry not included in the training data (Fig. 3K-O),
where DL-Sim produces particularly heterogeneous estimates
in the inclusion. In all cases, the naïve linear calibration model
systematically underestimated µa in both inclusions and in the
background material (Cal., Fig. 3A, F, K). Fluence correction
with Monte Carlo simulations yields an overestimation of
µa for the larger inclusion and towards the centre of the
background material, creating distorted edges of the inclusions
(GT-ϖ, Fig. 3B, G, L).
The estimation of µa from DL-Sim shows a systematic
overestimation of the µa in the background material at the
upper rim of the phantom, presumably caused by incor-
rect assumptions present in the 2D acoustic forward model
(DL-Sim, Fig. 3C, H, M). DL-Sim also shows artefacts, with
estimation of a high µa outside of the phantom area in the cou-
pling medium. In contrast, DL-Exp predicts spatially smooth
estimates with sharp borders around the phantom and inclu-
sions (DL-Exp, Fig. 3D, I, N). DL-Exp also shows improved
quantitative performance in the inclusions compared to DL-
Sim, across all inclusions (Fig. 3E, J, O). The same findings
are seen consistently in line profiles across the test phantom
set, with DL-Exp consistently outperforming all methods in
spatial accuracy and quantitative estimation, highlighting the
substantial added value of training on experimental data.

B. Absorption Coefficient Estimates at Depth Can Be

Fig. 4. PA signal calibration against reference optical absorption
Recovered With Deep Learning with a linear mapping function fails at depth. Linear mapping derived
We next analysed the quantitative performance of the from the training set was applied to the test set signals from the back-
ground (A) and inclusions (B). (C) Experimental and simulated PA signals
models in more detail. Cal. performs extremely well in the are shown as function of depth normalised to their maximum signal
background material at the surface of the phantoms (R=0.88), intensity. The shaded areas correspond to the standard deviation of the
providing an accurate estimate of µa independently of µs signal intensities. Vertical dashed lines highlight depths of approximately
1 mm and 13 mm, for which (D) and (E) show the scatter plot and linear
(Fig. 4A). This is not true for signals arising from deeper fit of all data points across all phantoms in the test set found at those
within the phantoms. In the inclusions at depth within the respective depths.
phantoms, there was no correlation with µa (R=→0.22)
(Fig. 4B), as expected due to depth-dependent optical attenu-
ation and spectral colouring [11]. Having established the performance limits of the baseline
To understand the limit at which the model breaks down, model (Cal.), we then examined the performance of the U-Net
we analysed the 25 training phantoms without any inclu- estimations. Application of these maintained or improved the
sions, showing that the correlation between PAI signal and excellent linear mapping in the background material (Table. I
µa systematically decreases with depth (Fig. 4C), ranging Background), while recovering a linear relationship between
from R=0.98 at 1 mm depth to R=0.15 at 13 mm depth the absorption estimate in the inclusions and the ground truth
(Fig. 4D-E). Although we observe the same overall trend (Table. I Inclusions), showing a greater robustness to depth-
between the experiment and simulation (Fig. 4C) the exper- dependent effects. Correcting the signals with the Monte Carlo
imental results do deviate from the monotonic exponential fluence reference yielded the best quantitative result with a
decay seen in the simulation. A number of factors could median error of 22% across the entire dataset. Unfortunately,
contribute to this observation: (1) the optical forward model it shows a clear SNR-dependency in performance, with poorer
assumes a literature-derived anisotropy and the applicability of performance at depth, and is most challenging to apply in
the Henyey-Greenstein scattering function [41], which might living subjects due to the lack of knowledge of ground truth
differ in our phantom materials; (2) we do not model the noise optical properties.
present in the experimental data; and (3) we do not perform While both deep learning methods underperform com-
3D acoustic forward modelling, thus not accounting for out of pared to GT-ϖ (vs. DL-Exp p=0.012, vs. DL-Sim p=10→13 ),
plane pressure waves or ultrasound focusing. DL-Exp (median error of 29%) significantly (p=10→7 )
1220 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 43, NO. 3, MARCH 2024

TABLE I
C OMPARISON OF P ERFORMANCE OF THE D IFFERENT M ETHODS .
VALUES A RE S HOWN AS M EDIAN ± I NTERQUARTILE R ANGE /2. R =
P EARSON C ORRELATION C OEFFICIENT ; G CNR = G ENERALISED
C ONTRAST- TO -N OISE R ATIO

outperformed DL-Sim (median error of 37%). We also

observed a systematic underestimation of high absorption
coefficients for both deep learning models. Since this effect
is consistent for the simulated and experimental data set,
we assume that this effect is caused by the composition of
the dataset and a larger, more diverse data set would be
required to counteract this effect. When evaluating inclusions
with µa ≃ 2.5 cm→1 , a likely in vivo scenario, we find that DL-
Exp begins to outperform GT-ϖ with regard to the correlation
coefficient and gCNR (Table. I µa ≃ 2.5 cm→1 ), demonstrating
the future potential of the approach in living subjects.

C. Experimentally Trained Deep Learning Models Are

Transferable to a Blood Flow Phantom Set
We next applied the Cal., DL-Sim, and DL-Exp µa esti-
mation methods in an independent flow phantom experiment
(Fig. 5); the GT-ϖ method could not be applied as ground truth Fig. 5. DL-Sim fails to preserve the dynamic range of oxygenation
optical properties are unknown. response. A Cross-sectional slice through the phantom with the blood-
carrying tube and the spatial absorption profiles along the dashed line.
The flow phantom was distinct from the training and B Mean blood oxygen saturation (sO2 ) derived with linear unmixing from
test material, as it was fabricated from a different base multispectral µa estimates for each of the methods over time, with spectra
and included a tube carrying blood that was chemically at specific time points shown respectively in C and D. A reference for
the expected µa spectrum of blood (scaled down by a factor of 3 for
deoxygenated over time. We found that DL-Exp estimated visualisation purposes) at the time points is shown as a dashed black
a nearly two-fold higher µa value in the blood tube and line.
DL-Sim a nearly three-fold increased µa value compared
to the baseline method (Fig. 5A). Thus, both methods were
able to recover a value closer to the expected reference of the end of the experiment compared to the reference Cal.
about 4.3 cm→1 at 800 nm [40]. When computing sO2 from method.
the µa estimates, DL-Sim produced a very poor dynamic Comparison of the mean spectra at different time points
range of sO2 during the imaging time, ranging from 60% (Fig. 5C,D) with literature-derived reference spectra for the
to 45%, rendering any biological applications of the derived µa of haemoglobin [40], however, revealed that none of the
biomarker potentially useless (Fig. 5B). DL-Exp, on the other methods were able to compensate for the spectral colouring
hand, was able to reproduce sO2 at the start of the exper- induced by water absorption in the coupling media, clearly
iment and marginally increased the dynamic range towards visible from 850-900 nm.
GRÖHL et al.: MOVING BEYOND SIMULATION: DATA-DRIVEN QUANTITATIVE PHOTOACOUSTIC IMAGING 1221

excellent spatial performance in phantoms, the U-Net µa

estimates were not able to retain the high spatial frequencies
of the in vivo PA image, which also translates to the resulting
sO2 estimates (Fig. 6C,E), most likely due to the lack of high
spatial frequency features in the training set.
We computed the relative µa -ratio between the aorta and
the spine (Fig. 6G), to understand whether the methods could
compensate depth-dependent decreases in signal intensity.
Physiologically, we would expect a higher µa and oxygenation
in the aorta compared to the spine due to higher average blood
volume and presence of both arteries and veins in the spine.
While the Cal. estimate gave approximately the same µa for
both regions, the U-Net estimates result in a 2-3 fold higher
µa in the aorta compared to the spine. Consistent with the
artefacts observed in the test phantoms, DL-Sim estimated
high µa values in the upper rim of the mouse body ans in
the water background.
Application of linear unmixing revealed that the general
trends of high and low oxygenation in the segmented organs
were quite consistent across all methods. A notable exception
was the left kidney, where both U-Net estimations showed
relatively high oxygenation, whereas the calibrated signal
estimated relatively low oxygenation. Computing the arterial
sO2 in the aorta reveals an 8 ± 5 percentage points increase
(p=0.002) in blood oxygenation on the DL-Exp µa estimate
compared to the calibrated signal (Fig. 6H), though all models
still underestimate the expected arterial blood oxygenation
in an anaesthetised mouse of about 94%-98% [42], perhaps
relating to the generally lower dynamic range seen in the blood
phantom.

IV. D ISCUSSION
We developed a supervised deep learning method to address
the challenge of estimating optical absorption coefficients in
PAI enabled by a collection of carefully annotated phantoms.
The dataset consists of an unprecedented 102 tissue-mimicking
phantoms spanning an important biological range of tissue
optical properties. When assuming independence of images at
21 captured wavelengths from 700 nm to 900 nm, our dataset
features a total of 2142 unique PA images of targets with
distinct and known µa and µ↔s distributions. Each image
was subject to manual segmentation and quality assessment,
then assigned reference optical absorption and scattering coef-
ficients as determined by independent DIS measurements.
Fig. 6. Absorption estimation methods applied in living subjects. We used digital twins of the phantoms to simulate a synthetic
(A) Calibrated PA signal (Cal.) and (B) corresponding oxygenation (sO2 ).
C) µa estimates of DL-Sim and (D) corresponding sO2 . E) µa estimates replica data set to analyse the differences between using
of DL-Exp and (F) corresponding sO2 . G shows the relative estimated simulated and synthetic data sets for model training.
µa ratios between the aorta and spinal region over nine subjects and
H shows the distribution of sO2 values in the aorta. The expected sO2
of 96%±2% is shown as a green area. A. Discussion of Results
We show that erroneous model assumptions have an influ-
ence on the quantitative accuracy of the deep learning model
D. Oxygenation Maps Derived From Learned µa
and that training with experimentally acquired phantoms
Estimates Show More Realistic Arterial Oxygen
has advantages in both accuracy and generalisability of the
Saturation
approach. Our results indicate that tissue-mimicking phantoms
Finally, we tasked Cal., DL-Sim, and DL-Exp methods to can be used as reliable calibration objects and that using
estimate µa in vivo (Fig. 6A,C,E) and subsequently calculated experimental training data is beneficial compared to synthetic
sO2 from multispectral estimates (Fig. 6B,D,F). Despite their training data, indicating the existence of a simulation gap.
1222 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 43, NO. 3, MARCH 2024

At the same time, we confirm that linear calibration schemes experimental and simulated data are expected to be sensitive
fail as they are unable to account for the exponential decay to changes in ϑ of the imaged material, which could be reasons
of fluence with depth. We find that using the experimental for the remaining discrepancies in the flow phantom data set.
data set yields improved accuracy across multiple test cases Similarly for the in vivo application in a mouse, DL-Exp gave a
compared to simulated data sets, at least given the simulation much more realistic representation of the absorption difference
pipeline used in this work. between the aorta and spine, and resulted in a substantial
The deep learning models were initially applied to a test elevation of the estimated arterial oxygen saturation. The
phantom set and results were compared with the naïve cali- final results remained an under-estimate compared to the
bration model and fluence compensation using Monte Carlo expected ⇐ 94% but suggest a greater resilience to depth-
modelling (assuming known optical properties). All methods dependent attenuation.
could recover some correlation of the estimated µa coefficients The U-Nets produced estimates with sharp edges on
compared to the reference µa . A striking result of DL-Exp is the blood-background interface in the flow phantom, likely
the overall improvement in spatial accuracy of the predictions, because of the consistent nature of the spatial frequency
which is demonstrated by smooth estimates with sharply content of the images. Contrary to the promising spectral
defined borders between segmentation classes, quantified by performance, when applied to the mouse data, the U-Nets
consistently high gCNR values. On the other hand, we found were not able to recover the high spatial frequency content
a median relative quantification error of ↑ 29% when training of the image, suggesting a need to enrich the training data set
on experimentally acquired data and ↑ 37% when training on with inclusions that better reflect the complexity of the in vivo
simulated data. In comparison, fluence correction with a Monte imaging application in order to achieve high spatial accuracy
Carlo model using the reference absorption and scattering in the general case.
measurements (GT-ϖ), achieves a relative absorption estima-
tion error of ↑ 22%, but is less accurate in terms of spatial B. Experimental Limitations
estimation at depth and not applicable to settings without While training on experimental phantom measurements
known optical tissue properties. Furthermore, GT-ϖ systemati- shows promise to advance quantitative µa estimates in PAI,
cally overestimates the absorption in depth, indicating a model there are also limitations that must be addressed in future
mismatch in the optical or acoustic forward models. DL-Sim work. Considering first experimental limitations, we only use
systematically overestimates the background absorption at the titanium dioxide to tune µ↔s and nigrosin to tune µa in the
upper side of the phantoms, which could either indicate a phantoms. This is important to minimise the influence of the
mismatch in the light model, or an over/underestimation of molecule- and concentration-dependent Grüneisen parameter
the limited view artefact in the forward model compared to but might introduce a wavelength-dependent bias to the train-
the experimental data as well. ing data that may challenge our assumption of independence
When evaluating using only imaging targets with µa ≃ of images across wavelengths. We assumed images from
2.5 cm→1 , we find that DL-Exp begins to outperform GT-ϖ on different wavelengths to be independent, to maximise the
several reported measures. By extending the training data set, number of training images even though images at neighbouring
we hope to achieve significantly better quantification results. wavelengths will have a high correlation. While such an
Improvement of the deep learning model to mimic traditional assumption could potentially lead to the introduction of a
iterative schemes might also be a promising way forward to wavelength-dependent bias, our results do not show an indi-
further improve quantitative performance. cation for it, suggesting that the assumption – while not being
When moving to completely different phantom and exper- ideal – did not systematically affect the learned algorithms.
imental data, the U-Nets performed less well in terms of If the assumption does not hold, it could impact the quan-
spatial predictions. The significant difference in the spatial tification of multispectrally derived biomarkers, such as blood
distributions of chromophores between the training and test oxygenation sO2 .
data was made intentionally and highlights the limited gen- A second experimental limitation of the study is the
eralisability of the method to data far outside of the training assumption that DIS measurements can be taken as a ground
data distribution. For example, the µa estimate of blood in truth [46], where in fact they may introduce substantial uncer-
the flow phantom was significantly increased although it was tainty [47] for example from modelling errors by a priori
still nearly a factor of three too low compared to the expected assumptions for the sphere parameters, sphere corrections,
µa from literature. We did not include 3D acoustic forward or measurement error for sample thicknesses. Accurate quan-
modelling with consideration of the acoustic attenuation of tification of optical properties is highly challenging, with
the medium and the angular and frequency response of the variation of up to 30% and 40% for µa and µ↔s , respectively,
transducers or the Grüneisen parameter ϑ in our numerical when quantifying the same sample with different instru-
models. We did not conduct measurements for estimations of ments [48]. We examined the influence of our uncertainties and
ϑ for our fabricated materials, however, the large potential found a standard error of 2.5% in the thickness measurements,
discrepancy of ϑ between blood (0.152-0.226) [43], water which translates to approx. 5% error of the optical absorp-
(↑ 0.11) [44], and mineral oil (↑ 0.8) [45] could be one tion coefficient and simulated initial pressure distribution.
explanation for the low µa estimates on the flow phantom In conjunction with the 10%-20% error we determined for
data. While the acoustic modelling aspect is only relevant the DIS system based on cross-reference with time-resolved
when using simulated training data, models trained on both near-infrared spectroscopy [27], our reference optical material
GRÖHL et al.: MOVING BEYOND SIMULATION: DATA-DRIVEN QUANTITATIVE PHOTOACOUSTIC IMAGING 1223

properties are subject to a standard error in the order of 20%. computational modelling of the forward processes involved
In future work, these uncertainties could be incorporated into in PAI, we demonstrate that our phantoms can be used
model training by using Bayesian deep learning methods. as reliable calibration objects. Our experiments confirm the
limitations of linear calibration schemes in accounting for
C. Computational Limitations
depth-dependent fluence effects and show that a deep learn-
We introduced a number of approximations during the ing method exclusively trained on experimentally acquired
Monte Carlo and k-space simulations, such as 2D acoustic datasets can provide better optical property estimation of
modelling (motivated by the out-of-plane symmetry of our previously unseen test phantoms compared to simulated data
phantoms), the laser source geometry and beam profile, and sets. With demonstrations in different test scenarios, including
absolute homogeneity of the phantom materials (i.e. sound in living subjects, deep learning shows potential for real-world
speed and density). We used the delay-and-sum algorithm as applications. Despite remaining limitations in spectral and
the acoustic inverse operator, which is the most commonly spatial accuracy, our work represents an exciting advancement
used reconstruction algorithm even though it is known to towards the development of accurate and reliable quantitative
introduce artefacts to the reconstructed images for our imaging PAI data analysis.
setup [49]. We assume that a combination of these modelling
errors in conjunction with the presence of measurement noise
deep inside tissue is the reason for the relatively high µa quan- D ISCLOSURES
tification error for GT-ϖ. The influence of noise is highlighted For the purpose of open access, the authors have applied a
when visualising the absorption estimated within larger inclu- Creative Commons Attribution (CC BY) licence to any Author
sions with high µa , leading to an increasing overestimation Accepted Manuscript version arising from this submission.
of µa with depth and a wavelength-dependent performance The data to reproduce the findings of this study will be
of GT-ϖ. Using a network architecture that better mirrors the made available via https://doi.org/10.17863/CAM.96644
QPAI problem might increase the quantification performance, upon publication of this manuscript. The respective
either by using a multi-path network with dedicated loss code work is available on GitHub at https://github.com/
functions [24], or by including a model into the network and BohndiekLab/end_to_end_phantom_QPAT.
learning the update for an iterative optimisation scheme [15].
D. Future Directions R EFERENCES
When only analysing results for absorbers with µa ≃ [1] S. Manohar and M. Dantuma, “Current and future trends in photoacous-
2.5 cm→1 , our findings suggest that a learning-based approach tic breast imaging,” Photoacoustics, vol. 16, Dec. 2019, Art. no. 100134.
[2] X. L. Deán-Ben, A. Ozbek, and D. Razansky, “Volumetric real-time
has distinct advantages that might allow it to eventually outper- tracking of peripheral human vasculature with GPU-accelerated three-
form fluence correction. Learning-based models might learn to dimensional optoacoustic tomography,” IEEE Trans. Med. Imag., vol. 32,
correct for effects such as measurement noise, out-of-plane no. 11, pp. 2050–2055, Nov. 2013.
[3] S. Y. Chuah et al., “Structural and functional 3D mapping of skin
signal contributions or signal blurring caused by the point tumours with non-invasive multispectral optoacoustic tomography,” Skin
spread function of the measurement system. Unfortunately, our Res. Technol., vol. 23, no. 2, pp. 221–226, May 2017.
results also show that the proposed approach is far from being [4] J. Aguirre et al., “Precision assessment of label-free psoriasis biomarkers
with ultra-broadband optoacoustic mesoscopy,” Nature Biomed. Eng.,
usable in a clinical setting. To achieve this, we require: (1) vol. 1, no. 5, pp. 1–8, May 2017.
more and broader training data with a focus on more realistic [5] F. Knieling et al., “Multispectral optoacoustic tomography for assess-
spatial chromophore distributions; (2) highly optimised and ment of Crohn’s disease activity,” New England J. Med., vol. 376, no. 13,
accurate optical and 3D acoustic forward models; (3) realistic pp. 1292–1294, Mar. 2017.
[6] M. Xu and L. V. Wang, “Photoacoustic imaging in biomedicine,” Rev.
noise models; as well as (4) model-based acoustic inversion Sci. Instrum., vol. 77, no. 4, Apr. 2006, Art. no. 041101.
schemes to obtain more faithful reconstructions of the initial [7] P. Beard, “Biomedical photoacoustic imaging,” Interface Focus, vol. 1,
pressure distributions. Using a domain transfer approach, such no. 4, pp. 602–631, Aug. 2011.
[8] H. Jaffe and D. A. Berlincourt, “Piezoelectric transducer materials,”
as used by Li et al. [24] or Dreher et al. [22] to complement Proc. IEEE, vol. 53, no. 10, pp. 1372–1386, Oct. 1965.
the experimental data set used in this work with additional [9] J. M. Vaughan, The Fabry–Perot Interferometer: History, Theory, Prac-
simulated data might also be a promising direction. Finally, tice and Applications. Evanston, IL, USA: Routledge, 2017.
[10] B. Cox, J. Laufer, and P. Beard, “The challenges for quantitative pho-
in order to validate the trained methods, diverse in vivo toacoustic imaging,” in Proc. SPIE, vol. 7177, pp. 294–302, Feb. 2009.
test data sets have to be acquired that are representative of [11] B. Cox, J. G. Laufer, S. R. Arridge, and P. C. Beard, “Quantitative spec-
the target applications and annotated with high-quality and troscopic photoacoustic imaging: A review,” J. Biomed. Opt., vol. 17,
no. 6, 2012, Art. no. 061202.
high-confidence labels of the underlying physiology. Since [12] B. T. Cox, S. R. Arridge, K. P. Köstli, and P. C. Beard, “Two-dimensional
ground truth knowledge of the underlying optical properties quantitative photoacoustic image reconstruction of absorption distribu-
in vivo is rarely available, the validation of learned algorithms tions in scattering media by use of a simple iterative method,” Appl.
Opt., vol. 45, no. 8, pp. 1866–1875, 2006.
remains a major roadblock on the path towards quantitative [13] T. Tarvainen, A. Pulkkinen, B. T. Cox, J. P. Kaipio, and S. R. Arridge,
photoacoustics. “Bayesian image reconstruction in quantitative photoacoustic tomog-
raphy,” IEEE Trans. Med. Imag., vol. 32, no. 12, pp. 2287–2298,
V. C ONCLUSION Dec. 2013.
We introduce a collection of well-characterised phantoms to [14] J. Buchmann, B. Kaplan, S. Powell, S. Prohaska, and J. Laufer, “Quan-
titative PA tomography of high resolution 3-D images: Experimental
enable supervised training and evaluation of learned quantita- validation in a tissue phantom,” Photoacoustics, vol. 17, Mar. 2020,
tive PAI methods. Leveraging the power of deep learning and Art. no. 100157.
1224 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 43, NO. 3, MARCH 2024

[15] A. Hauptmann et al., “Model-based learning for accelerated, limited- [33] I. Wolf et al., “The medical imaging interaction toolkit (MITK): A toolkit
view 3-D photoacoustic tomography,” IEEE Trans. Med. Imag., vol. 37, facilitating the creation of interactive software by extending VTK and
no. 6, pp. 1382–1393, Jun. 2018. ITK,” Proc. SPIE, vol. 5367, pp. 16–27, May 2004.
[16] S. Antholzer, M. Haltmeier, and J. Schwab, “Deep learning for pho- [34] Q. Fang and D. A. Boas, “Monte Carlo simulation of photon migration
toacoustic tomography from sparse data,” Inverse Problems Sci. Eng., in 3D turbid media accelerated by graphics processing units,” Opt. Exp.,
vol. 27, no. 7, pp. 987–1005, Jul. 2019. vol. 17, no. 22, pp. 20178–20190, 2009.
[17] C. Cai, K. Deng, C. Ma, and J. Luo, “End-to-end deep neural network [35] B. E. Treeby and B. T. Cox, “k-wave: MATLAB toolbox for the
for optical inversion in quantitative photoacoustic imaging,” Opt. Lett., simulation and reconstruction of photoacoustic wave fields,” Proc. SPIE,
vol. 43, no. 12, pp. 2752–2755, Jun. 2018. vol. 15, no. 2, 2010, Art. no. 021314.
[18] J. Gröhl, T. Kirchner, T. Adler, and L. Maier-Hein, “Confidence estima- [36] G. M. Hale and M. R. Querry, “Optical constants of water in the 200-nm
tion for machine learning-based quantitative photoacoustics,” J. Imag., to 200-µm wavelength region,” Appl. Opt., vol. 12, no. 3, pp. 555–563,
vol. 4, no. 12, p. 147, Dec. 2018. 1973.
[19] T. Chen, T. Lu, S. Song, S. Miao, F. Gao, and J. Li, “A deep learning [37] J. Gröhl et al., “SIMPA: An open-source toolkit for simulation and image
method based on U-Net for quantitative photoacoustic imaging,” Proc. processing for photonics and acoustics,” Proc. SPIE, vol. 27, no. 8,
SPIE, vol. 11240, pp. 216–223, Feb. 2020. Apr. 2022, Art. no. 083010.
[20] J. Gröhl, M. Schellenberg, K. Dreher, and L. Maier-Hein, “Deep learn- [38] M. Gehrung, S. E. Bohndiek, and J. Brunker, “Development of a blood
ing for biomedical photoacoustic imaging: A review,” Photoacoustics, oxygenation phantom for photoacoustic tomography combined with
vol. 22, Jun. 2021, Art. no. 100241. online pO2 detection and flow spectrometry,” Proc. SPIE, vol. 24, no. 12,
[21] M. Schellenberg et al., “Photoacoustic image synthesis with gen- 2019, Art. no. 121908.
erative adversarial networks,” Photoacoustics, vol. 28, Dec. 2022, [39] A. Rodriguez-Molares et al., “The generalized contrast-to-noise ratio:
Art. no. 100402. A formal definition for lesion detectability,” IEEE Trans. Ultrason.,
[22] K. K. Dreher et al., “Unsupervised domain transfer with conditional Ferroelectr., Freq. Control, vol. 67, no. 4, pp. 745–759, Apr. 2020.
invertible neural networks,” 2023, arXiv:2303.10191. [40] S. L. Jacques, “Optical properties of biological tissues: A review,” Phys.
[23] J. Gröhl et al., “Learned spectral decoloring enables photoacoustic Med. Biol., vol. 58, no. 11, pp. R37–R61, Jun. 2013.
oximetry,” Sci. Rep., vol. 11, no. 1, p. 6565, Mar. 2021. [41] L. C. Henyey and J. L. Greenstein, “Diffuse radiation in the galaxy,”
[24] J. Li et al., “Deep learning-based quantitative optoacoustic tomography Astrophys. J., vol. 93, pp. 70–83, Jan. 1941.
of deep tissues in the absence of labeled experimental data,” Optica, [42] A. M. Loeven, C. N. Receno, C. M. Cunningham, and L. R. DeRuisseau,
vol. 9, no. 1, pp. 32–41, 2022. “Arterial blood sampling in male CD-1 and C57BL/6J mice with 1%
[25] L. Hacker et al., “A copolymer-in-oil tissue-mimicking material with isoflurane is similar to awake mice,” J. Appl. Physiol., vol. 125, no. 6,
tuneable acoustic and optical characteristics for photoacoustic imaging pp. 1749–1759, Dec. 2018.
phantoms,” IEEE Trans. Med. Imag., vol. 40, no. 12, pp. 3593–3603, [43] D.-K. Yao, C. Zhang, K. Maslov, and L. V. Wang, “Photoacoustic
Dec. 2021. measurement of the Grüneisen parameter of tissue,” Proc. SPIE, vol. 19,
[26] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional net- no. 1, 2014, Art. no. 017007.
works for biomedical image segmentation,” in Proc. Int. Conf. Med. [44] E. Petrova, S. Ermilov, R. Su, V. Nadvoretskiy, A. Conjusteau, and
Image Comput. Comput.-Assist. Intervent. Cham, Switzerland: Springer, A. Oraevsky, “Using optoacoustic imaging for measuring the tem-
2015, pp. 234–241. perature dependence of Grüneisen parameter in optically absorbing
[27] L. Hacker, “Validation of photoacoustic imaging biomarkers in cancer solutions,” Opt. Exp., vol. 21, no. 21, pp. 25077–25090, 2013.
biology,” Ph.D. dissertation, Univ. Cambridge, Cambridge, U.K., 2022. [45] S. Liang, B. Lashkari, S. S. S. Choi, V. Ntziachristos, and A.
[28] J. W. Pickering, S. A. Prahl, N. Van Wieringen, J. F. Beek, Mandelis, “The application of frequency-domain photoacoustics to
H. J. Sterenborg, and M. J. Van Gemert, “Double-integrating-sphere temperature-dependent measurements of the Grüneisen parameter in
system for measuring the optical properties of tissue,” Appl. Opt., lipids,” Photoacoustics, vol. 11, pp. 56–64, Sep. 2018.
vol. 32, no. 4, pp. 399–410, 1993. [46] E. Zamora-Rojas, B. Aernouts, A. Garrido-Varo, D. Pérez-Marín,
[29] S. A. Prahl, “Everything I think you should know about inverse adding- J. E. Guerrero-Ginel, and W. Saeys, “Double integrating sphere mea-
doubling,” Oregon Med. Laser Center, St. Vincent Hospital, Portland, surements for estimating optical properties of pig subcutaneous adipose
OR, USA, Tech. Rep., 2011, pp. 1–74. tissue,” Innov. Food Sci. Emerg. Technol., vol. 19, pp. 218–226,
[30] C. J. M. Jones and P. R. T. Munro, “Stability of gel wax based optical Jul. 2013.
scattering phantoms,” Biomed. Opt. Exp., vol. 9, no. 8, pp. 3495–3502, [47] J. F. Beek et al., “In vitro double-integrating-sphere optical properties
Aug. 2018. of tissues between 630 and 1064 nm,” Phys. Med. Biol., vol. 42, no. 11,
[31] J. Joseph, M. R. Tomaszewski, I. Quiros-Gonzalez, J. Weber, J. Brunker, pp. 2255–2261, Nov. 1997.
and S. E. Bohndiek, “Evaluation of precision in optoacoustic tomogra- [48] A. Pifferi et al., “Performance assessment of photon migration instru-
phy for preclinical imaging in living subjects,” J. Nucl. Med., vol. 58, ments: The MEDPHOT protocol,” Appl. Opt., vol. 44, no. 11,
no. 5, pp. 807–814, May 2017. pp. 2104–2114, 2005.
[32] T. Else, J. Gröhl, L. Hacker, and S. E. Bohndiek, “PATATO: A Python [49] M. Xu and L. V. Wang, “Universal back-projection algorithm for
photoacoustic tomography analysis toolkit,” Proc. SPIE, vol. PC12379, photoacoustic computed tomography,” Phys. Rev. E, Stat. Phys. Plasmas
Mar. 2023, Art. no. PC123790T. Fluids Relat. Interdiscip. Top., vol. 71, no. 1, Jan. 2005, Art. no. 016706.