Development of a blood oxygenation phantom for photoacoustic tomography combined with online pO2 detection and flow spectrometry

TL;DR

This paper presents a low-cost flow phantom system for validating photoacoustic tomography measurements of blood oxygen saturation, enabling dynamic assessment and correction of measurement inaccuracies in a controlled setting.

Contribution

The authors developed a novel, low-cost flow phantom with online sensors for validating and improving photoacoustic sO2 measurements in vitro.

Findings

Demonstrated the phantom's ability to evaluate sO2 using blood oxygenation changes.

Showed the influence of multispectral processing on sO2 accuracy.

Validated the phantom as a tool for future calibration and parameter studies.

Abstract

Photoacoustic tomography (PAT) is intrinsically sensitive to blood oxygen saturation (sO2) in vivo. However, making accurate sO2 measurements without knowledge of tissue- and instrumentation-related correction factors is extremely challenging. We have developed a low-cost flow phantom system to facilitate validation of photoacoustic tomography systems. The phantom is composed of a flow circuit, which is partially embedded within a tissue mimicking phantom, with independent sensors providing online monitoring of the optical absorption spectrum and partial pressure of oxygen in the tube. We first establish the flow phantom using two small molecule dyes that are frequently used for photoacoustic imaging: methylene blue (MB) and indocyanine green (ICG). We then demonstrate the potential of the phantom for evaluating sO2 using chemical oxygenation and deoxygenation of blood in the phantom. Using this dynamic assessment of the photoacoustic sO2 measurement in phantoms in relation to a ground truth, we explore the influence of multispectral processing and spectral coloring on accurate assessment of sO2. Future studies could exploit this low-cost dynamic flow phantom to validate fluence correction algorithms and explore additional blood parameters such as pH, and also absorptive and other properties of different fluids.

Figures

Equations

$$H(\mathbf{r},\lambda )={\mu }_a(\mathbf{r},\lambda )\Phi [\mathbf{r},\lambda ,{\mu }_a(\mathbf{r},\lambda ),{\mu }_s^{\prime }(\mathbf{r},\lambda )].$$

$${\text{sO}}_2(\%)=\frac{100}{{\left[23400\times ({\text{pO}}_2{)}^3+150\times ({\text{pO}}_2)\right]}^{-1}+1},$$

$$\frac{dQ(t)}{dt}=F\left[c-\frac{Q(t)}{V}\right].$$

$$Q(0)=Q_0=0;$$

$$\frac{Q(t)}{V}=c\left[1-e^{-\frac{F}{V}t}\right].$$

$$\text{x}=\text{SC}+n,$$

$$\text{Cˆ}=\underset{C}{\mathrm{argmin}}{\Vert \text{SC}-\text{x}\Vert }_2^2,$$

$$\text{Cˆ}={\text{xS}}^{+},$$

Full text

***Joanna Brunker, \linkablejb2014@cam.ac.uk **

1 Introduction

Photoacoustic tomography (PAT) exploits optically generated ultrasound to provide images that combine the high contrast and spectral specificity of optical imaging with the high spatial resolution of ultrasound. In particular, PAT has been widely used to image blood hemoglobin concentration and oxygenation, which have the potential to inform on a range of pathophysiologies, from tumour aggressiveness [1] and treatment response [2, 3] to intestinal inflammation associated with Crohn’s disease [4] and colitis [5]. The derivation of such images is usually based on spectral unmixing to resolve the differential absorption contributions of oxy- (HbO2) and deoxy-hemoglobin (Hb). Total hemoglobin concentration (THb) is typically taken as the sum of the contributions to the photoacoustic signal $P(\mathbf{r},\lambda)$ from HbO2 and Hb, while hemoglobin oxygenation (sO2) is taken as the ratio of HbO2 to THb.

Unfortunately, the assessment of HbO2 and Hb content from photoacoustic data is not trivial. Estimation of these chromophore concentrations from images taken at multiple wavelengths is commonly achieved using least-squares fitting of reference HbO2 and Hb spectra [6], which are tabulated in the literature [7, 8, 9]; however, these spectra are recorded under in vitro conditions thus can vary substantially from those experienced within an in vivo study. Moreover, the measured photoacoustic signal $P(\mathbf{r},\lambda)$ is not directly proportional to the absorbed energy density $H(\mathbf{r},\lambda)$. $H(\mathbf{r},\lambda)$ is the product of the light fluence $\Phi$, which itself varies as a function of $\mathbf{r}$ and $\lambda$, as well as $\mu_{a}(\mathbf{r},\lambda)$ and the reduced scattering coefficient $\mu^{\prime}_{s}(\mathbf{r},\lambda)$:

[TABLE]

This co-dependence of $H(\mathbf{r},\lambda)$ on both absorption and light fluence leads to an effect known as “spectral coloring”, where variations in local fluence bias the measured optical absorption distribution [10, 11]. Despite this challenge, the amplitude and temporal profiles of photoacoustic waves have been shown to be linearly dependent on the blood sO2 in vitro [12]. In theory, only two wavelengths are necessary to calculate two concentrations, but practical limitations mean the choice and number of wavelengths becomes critical for accurate sO2 measurements [13]. Alternative approaches beyond the standard linear unmixing model have been shown to improve sO2 measurement accuracy, in particular by accounting for the absorbed energy distribution through: internal irradiation [14]; diffusion theory modelling [15, 16, 17, 18, 19]; Monte Carlo simulations [20]; model-based iterative minimisation [21, 22, 23]; and linear superposition of reference fluence base spectra [24].

Several studies have explored ways to validate photoacoustic images using phantoms with well-characterised optical and acoustic properties [25, 26]. Validation of photoacoustically measured sO2 is possible using CO-oximetry [12, 15, 27], pulse oximetry [28] and blood-gas analysis [29]; correlation between pO2 and sO2 has also been used to study oxygen-hemoglobin binding [30]. However, most studies have worked with static blood samples with limited control on the sO2. Allowing blood to flow in a circuit [27] and with the ability to vary the sO2 provides a more versatile platform for investigating PAT oxygenation measurements.

In this work, we created a low-cost flow phantom system with online monitoring to facilitate validation of PAT systems. The phantom was tested by circulating different concentrations of methylene blue (MB) and indocyanine green (ICG). The phantom was then applied to explore the accuracy of sO2 assessment using PAT, with sO2 values ranging from $0\text{\,}\mathrm{\char 37\relax}$ to $100\text{\,}\mathrm{\char 37\relax}$. We studied the impact of evaluating sO2 using different reference spectra for HbO2 and Hb, and also the effect of spectral coloring, showing discrepancies of up to $60\text{\,}\mathrm{\char 37\relax}$ between the actual and measured sO2. Our findings highlight the importance of careful choice of spectra for unmixing and the development of fluence correction models to improve the biological relevance of sO2 measurements derived from PAT images.

2 Methods

2.1 Flow system

The flow system (Fig. 1) enables fluids to be circulated within a vessel-mimicking tube embedded in a tissue-mimicking agar phantom placed in the chamber of a photoacoustic imaging system. Online (and offline) flow spectrometry and pO2 detection provide independent validation of the spectral measurements made by the imaging system. Details of the tissue-mimicking phantom, the photoacoustic imaging system, the spectrometric validation, the pO2 detection, and the spectral analysis are described in the following subsections.

2.1.1 Tissue-mimicking phantom

The tissue-mimicking phantom used for photoacoustic imaging comprised either static or flowing fluids within a tube (“vessel”) embedded within an agar (“tissue”) cylinder, $19\text{\,}\mathrm{mm}$ in diameter. Agar phantoms were prepared by heating a solution of 1.5% w/v agar (05039, Fluka) in water, and then adding 2.1% v/v Intralipid (I141, Sigma-Aldrich), before pouring into an open $20\text{\,}\mathrm{mL}$ syringe (with the injection end removed), in the centre of which was positioned a needle supporting the tubing vertically. The solidified agar phantom had a scattering coefficient of $5\text{\,}{\mathrm{cm}}^{-1}$ due to the Intralipid, but the absorption coefficient was assumed to be negligible.

To provide optical absorption for assessment of spectral coloring, nigrosin (198285, Sigma-Aldrich) was also added to the tissue-mimicking phantom formulation to produce phantoms with an absorption coefficient of $0.05\text{\,}{\mathrm{cm}}^{-1}$ or $0.1\text{\,}{\mathrm{cm}}^{-1}$ at $564\text{\,}\mathrm{nm}$ (the peak of the nigrosin spectrum).

2.1.2 Photoacoustic tomography (PAT)

For photoacoustic imaging, a small animal imaging system (MultiSpectral Optoacoustic Tomography (MSOT) inVision 256-TF, iThera Medical) was used. Briefly, a tunable optical parametric oscillator (OPO) pumped by an Nd:YAG laser provides excitation pulses with a duration of $9\text{\,}\mathrm{ns}$ at wavelengths from $660\text{\,}\mathrm{nm}$ to $1200\text{\,}\mathrm{nm}$ at a repetition rate of $10\text{\,}\mathrm{Hz}$ with a wavelength tuning speed of $10\text{\,}\mathrm{ms}$ and a peak pulse energy of $90\text{\,}\mathrm{mJ}$ at $720\text{\,}\mathrm{nm}$. Ten arms of a fibre bundle provide near-uniform illumination over a disk extending approximately $8\text{\,}\mathrm{mm}$ along the imaging chamber. Photoacoustic signals are detected using 256 toroidally focused ultrasound transducers with a centre frequency of $5\text{\,}\mathrm{MHz}$ (60% bandwidth), organized in a concave array of $270\text{\,}\mathrm{\SIUnitSymbolDegree}$ angular coverage and a radius of curvature of $4\text{\,}\mathrm{cm}$.

2.1.3 Detection of pO2

Two oxygen fluorescence quenching needle probes (NX-BF/O/E, Oxford Optronix) were inserted into the flow circuit before and after the tissue-mimicking phantom. A touch-screen monitor (OxyLite Pro, Oxford Optronix) displayed the temperature and partial pressure of oxygen (pO2) real-time, and these data were downloaded via an Arduino UNO and read in MATLAB. The relationship between the oxygen saturation (sO2) and the partial pressure (pO2) in blood is described by the characteristic sigmoid-shaped oxygen-hemoglobin dissociation curve. A widely accepted fit to this curve is given by the Severinghaus equation [31, 32]:

[TABLE]

which was used to convert our pO2 measurements into sO2.

2.1.4 Online flow spectrometry

Absorption spectra were recorded continuously via AvaSoft software using a light source (Avalight-HAL-S-Mini, Avantes) and spectrometer (AvaSpec-ULS2048-USB2-VA-50, Avantes) as the fluid passed through a flow cell (170700-0.5-40, Hellma Analytics).

2.1.5 Offline spectrometry

To independently validate spectrophotometric measurements made in the flow circuit, optical absorbance spectra were also recorded using a microplate spectrometer (CLARIOstar, BMG LABTECH). Fluid samples were measured in a 96-well plate (Corning Costar).

2.2 Flow system characterization

2.2.1 Tubing assessment

The optimum tubing was determined by comparing photoacoustic images obtained of agar phantoms containing various tube types filled with a $25\text{\,}\mathrm{\SIUnitSymbolMicro mol}$ ICG solution (Sigma-Aldrich I2633). The tubes were labelled according to their nominal inner and outer diameters (I.D / O.D.) in $\text{\,}\mathrm{\SIUnitSymbolMicro m}$, but were also made of different materials: polypropene 2660/2800 (Alliance Online PSTS0007); THV500 2800/3150, 500/600 (Paradigm Optics); silicone 1570/2410, 630/1190, 300/630 (VWR 228-0256, 228-0254, 228-0253); PVC 1500/2100 (VWR 228-3857); PMMA 667/1000, 432/865, 375/500 (Paradigm Optics); polythene 580/960 (Smiths Medical 800/100/200 12665497).

2.2.2 Dye dilution series

A dilution series was used to demonstrate the utility of the flow circuit for injecting different fluids into the closed PAT system, allowing PAT to be performed concurrently with online spectrometry. Methylene blue (MB, 50484, Fluka) and indocyanine green (ICG, I2633, Sigma-Aldrich) solutions were prepared by diluting concentrations of $500\text{\,}\mathrm{\SIUnitSymbolMicro M}$ and $100\text{\,}\mathrm{\SIUnitSymbolMicro M}$ (for MB and ICG respectively) in deionized water. First, four spectra were measured and averaged for samples of each concentration placed in a 96-well plate in the CLARIOstar spectrophotometer. The concentrations were then flushed individually through the flow system, starting with deionized water and then sequentially with increasing concentration. Ten online spectra were recorded over a range of $333\text{\,}\mathrm{nm}$ to $1100\text{\,}\mathrm{nm}$, and ten single-slice PA images (no pulse-to-pulse averaging) were acquired for 17 wavelengths ($660\text{\,}\mathrm{nm}$, $664\text{\,}\mathrm{nm}$, $680\text{\,}\mathrm{nm}$, $684\text{\,}\mathrm{nm}$, $694\text{\,}\mathrm{nm}$, $700\text{\,}\mathrm{nm}$, $708\text{\,}\mathrm{nm}$, $715\text{\,}\mathrm{nm}$, $730\text{\,}\mathrm{nm}$, $735\text{\,}\mathrm{nm}$, $760\text{\,}\mathrm{nm}$, $770\text{\,}\mathrm{nm}$, $775\text{\,}\mathrm{nm}$, $779\text{\,}\mathrm{nm}$, $800\text{\,}\mathrm{nm}$, $850\text{\,}\mathrm{nm}$, $950\text{\,}\mathrm{nm}$), taking the mean of the ten single-slice images for each concentration. The circuit was flushed with water between each concentration.

2.2.3 Dynamic concentration change

A dynamic concentration change was used to illustrate the possibility for real-time spectroscopic and photoacoustic measurements. Continuous acquisition of online spectra and single-slice photoacoustic images was commenced once deionized water was circulating within the flow system. After a certain time, a high concentration dye solution (either $500\text{\,}\mathrm{\SIUnitSymbolMicro M}$ MB or $100\text{\,}\mathrm{\SIUnitSymbolMicro M}$ ICG) was injected using the syringe pump at ($100\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{L}\mathrm{/}\min$).

The amount (moles) of the dye $Q(t)$ in the circuit at any given time $t$ can be modelled using a first order differential equation. In Eq. (3), the rate of change $dQ(t)/dt$ is equal to the difference between the inflow and outflow amounts, expressed in terms of the flow rate $F$ induced by the syringe pump ($100\text{\,}\mathrm{\SIUnitSymbolMicro}\mathrm{L}\mathrm{/}\min$), the concentration $c$ of dye injected ($500\text{\,}\mathrm{\SIUnitSymbolMicro M}$ or $100\text{\,}\mathrm{\SIUnitSymbolMicro M}$) and the volume $V$ ($5\text{\,}\mathrm{mL}$) of circulating fluid:

[TABLE]

The initial condition is given by Eq. (4), giving rise to Eq. (5), which is the solution for the dye concentration $Q/V$ as a function of time. In the limit $t\rightarrow\infty$, $Q(t)/V\rightarrow c$, as expected.

[TABLE]

[TABLE]

2.3 Blood oxygenation measurements

Mouse blood, predominantly made up of strains B6 and 129SvEv, was collected post mortem from the animal facility at the Cancer Research UK Cambridge Institute, complying with the UK Animals (Scientific Procedures) Act 1986. Each mouse contributed about $1\text{\,}\mathrm{mL}$ to a pool of about $5\text{\,}\mathrm{mL}$ blood required for the flow circuit. Chemicals were added to preserve, oxygenate and deoxygenate the blood. Prior to the experiment, ethylenediaminetetraacetic acid (EDTA) anti-coagulant (9002-07-7, Sigma-Aldrich) was added to the fresh blood, which was kept in a refrigerator under $4\text{\,}\mathrm{\SIUnitSymbolCelsius}$ for no more than 72 hours. During the experiment, the blood was returned to room temperature, and the oxygenation was controlled chemically by adding $0.2\text{\,}\mathrm{\char 37\relax}$ v/v hydrogen peroxide, H2O2 (7722-84-1, Sigma-Aldrich), for oxygenation [33] and sodium hydrosulfite (7775-14-6, ACROS Organics) for deoxygenation [34]. The sodium hydrosulfite was dissolved in phosphate-buffered saline (PBS) $\sim 0.03\text{\,}\mathrm{\char 37\relax}$ w/v for injection into the flow circuit.

2.4 Spectral and statistical analysis

PAT contrast is provided by optical absorbers within the field of illumination and detection. In general, PAT does not have sufficient resolution to visualize individual molecules and therefore each image pixel (or voxel) corresponds to more than one optical absorber, and a “spectral unmixing” approach is required in order to extract the individual spectral components. The linear mixture model [35] assumes that the measured spectrum is a linear combination of distinct spectra:

[TABLE]

where x is the $M$-wavelength $\times$ $N$-pixels measurement matrix, S is the $M\times K$ matrix of $K$ spectra (“end-members”), C is the unknown $K\times N$ matrix of end-member abundances (concentrations), and $n$ is measurement noise. Neglecting $n$, an estimation Ĉ of the absorber concentrations can be calculated from Eq. (6) by solving the following least-squares problem [11, 36]:

[TABLE]

with the solution

[TABLE]

where $\textbf{S}^{+}$ is the pseudoinverse of S. The performance of spectral unmixing carried out on acquired PA images was compared using three different spectra: those from the literature [7, 8, 9], and those experimentally measured in this study from either the online spectrometer cell or the offline plate reader. In addition, the online flow spectra were used to provide an independent ground truth for sO2; this entailed unmixing of the online flow spectra with the spectra for oxy- and deoxy- hemoglobin also measured live during the experiments.

To correct for the spectrally varying fluence (“spectral coloring”) when using optically absorbing phantoms, the image intensities in the tube region were divided by the nigrosin absorption spectrum. Specifically, the nigrosin absorption spectrum was normalised to the known absorption coefficient (at the $564\text{\,}\mathrm{nm}$ peak) and re-sampled at the 17 experimental wavelengths; the multispectral PA data were then divided by a wavelength-specific factor calculated from exponential decay over the $8.5\text{\,}\mathrm{mm}$ distance (the background agar material impregnated with absorbing nigrosin) from the outer edge of the phantom to the tube wall. All data and source code used in this publication is available on https://doi.org/10.17863/CAM.40365 and GitHub (https://github.com/9xg/flow-phantom).

3 Results

3.1 Flow system characterization

Eleven different tubes were assessed for their photoacoustic imaging suitability (Fig. 2). Tubes made from PMMA (I.D./O.D. 375/500, 432/865, 667/1000) showed the lowest signal-to-background (SBR) ratios ranging from 0.15 to 0.37. The other tubes materials show a clear trend of increasing SBR with larger inner diameters. Silicone tubes present high SBRs, even for low I.D./O.D. ratios (300/600, 630/1190, 1570/2410). Three PVC tubes showed very similar performance (I.D./O.D. 1500/2100, 1570/2410, and 2660/2800). The PVC tube with 1500/2100 presented a uniform, circular appearance in images with high SBR. It was therefore selected for the remaining experiments.

Dilutions of MB and ICG were tested inside the flow phantom under closed conditions i.e. with a steady concentration flowing in the circuit. By calculating the mean PA signal intensities over a manually segmented tube cross-sectional region of interest (ROI) for 17 different wavelengths it was possible to compare the PA spectra with those acquired during online flow spectrometry (Fig. 3 (a-d)). Unexpectedly, the two independent spectral measurements show poor agreement in terms of relative intensities and spectral shape for the different dye concentrations. In particular, it is notable that the PA signal intensities (Fig. 3b) show a more pronounced shift in spectral peak with increasing ICG concentration compared with the online spectrometer data (Fig. 3d); this may be a result of spectral coloring where light attenuation at the $800\text{\,}\mathrm{nm}$ peak suppresses the PA signal intensity deeper within the tube, leading to an overall reduction in mean intensity. For MB the discrepancies are difficult to discern seeing as the prominent spectral features occur below the PA imaging range, but it is interesting that the relative PA intensities are about half those for ICG suggesting that MB has poorer PA signal generation efficiency. Overall, these spectral inconsistencies raise a question about which are the most suitable endmember spectra for unmixing PAT images of MB and ICG.

The spectra measured with the online spectrometer and offline in the plate reader were used to perform concentration-specific unmixing of the mean ROI intensities, and these unmixed intensities were compared with those calculated using the literature spectra incorporated within the PA analysis software. These literature spectra are shown in Fig. 4 (a-b) along with online and offline spectra for example concentrations (Fig. 4 (c-f)).

Figure 5 (a-b) shows a linear relationship between dye concentration and spectrally unmixed PA intensities up to $100\text{\,}\mathrm{\SIUnitSymbolMicro mol}$ for MB, and $40\text{\,}\mathrm{\SIUnitSymbolMicro mol}$ for ICG, respectively. Divergence between the two types of unmixing is likely due to concentration dependent changes in the spectra of these dyes, which are particularly prominent for ICG.

Having explored the application of the flow circuit under static concentration values, we then examined the response of the PAT instrument to dynamic changes in dye concentrations for MB and ICG (Fig. 5 (c-d)). PA images were unmixed using the literature spectra and those recorded live using the online spectrometer, and then converted to absolute concentration values using the linear models calculated and plotted in (a-b). Neglecting the initial lag phase (which could not be experimentally determined and therefore was not incorporated into the model), the rates of concentration change match reasonably well with those predicted by the model (Eq. 5). However, there is a notable discrepancy between the dynamics calculated using the literature and online spectra for ICG unmixing; if the lag phase were corrected for, the online unmixing would be the closer match to the model and this is again likely to be a consequence of the concentration-dependent change in spectral shape, which is incorporated in the online spectra but not the single literature spectrum.

3.2 Blood oxygenation measurements

To demonstrate the closed nature of the blood flow circuit, blood was fully oxygenated by adding approximately $15\text{\,}\mathrm{\SIUnitSymbolMicro L}$ $0.2\text{\,}\mathrm{\char 37\relax}$ v/v H2O2 to $8\text{\,}\mathrm{mL}$ mouse blood and then injecting this into the flow system and circulating it for several minutes. Figure 6 shows the blood oxygen saturation (sO2) measured within the flow circuit by: the pO2 probe; the online spectrometer and PAT. The sO2 was calculated from the pO2 data using Equation 2 and from the online spectrometer and PAT data through spectral unmixing. To enable comparison between the methods independently, unmixing was performed using the spectra inherent to each system: for the online spectrometer, experimentally measured spectra were used as endmembers, while for PAT data, the literature spectra were used. As hoped, all three sO2 measurements show consistent values over time. However, the sO2 calculated from the unmixed PA images under-reads by about $13\text{\,}\mathrm{\char 37\relax}$ compared to the ground truth sO2 calculated from the pO2 probe and the online flow spectrometer.

In addition to investigating static blood oxygenation within the circuit, the full range of blood sO2 values ($100\text{\,}\mathrm{\char 37\relax}$ to $0\text{\,}\mathrm{\char 37\relax}$) were explored by injecting sodium hydrosulfite into the flow circuit in order to gradually deoxygenate the blood. The blood sO2 values cover the expected dynamic range when calculated from both the pO2 probe and the online spectrometer (Fig. 7a); however, PA signal intensities show markedly different behaviors when unmixed with spectra (Fig. 7b) from the literature or from the online flow spectrometer. The sO2 values calculated by unmixing with the experimentally measured spectra (red line, Fig. 7a) show good agreement with the spectrometer-derived ground truth values (yellow dotted line, Fig. 7a), except for somewhat over-reading the oxygenation for sO2 values below about $20\text{\,}\mathrm{\char 37\relax}$. They also show a dynamic range comparable to that obtained from the pO2 probe. However, the sO2 values calculated by unmixing with the literature spectra surprisingly exhibit a dramatically reduced dynamic range ($86\text{\,}\mathrm{\char 37\relax}$ to $58\text{\,}\mathrm{\char 37\relax}$); this may be due to the unmixing spectra and/or spectral coloring.

Two further investigations explored the potential impact of spectral colouring. First, sO2 measurements across the tube diameter were calculated (Fig. 8a). Unlike the results in Fig. 6 and Fig. 7 which were calculated from the mean of the ROI defined on the tube cross-sectional area, the results in Fig. 8a entailed radial unmixing using the mean of pixels found in a circle around the perimeter of the ROI, and repeating this for circle diameters reducing in one pixel ($75\text{\,}\mathrm{\SIUnitSymbolMicro m}$) increments. For the fully oxygenated blood the sO2 values are consistent across the tube diameter, but as the blood becomes deoxygenated there is a clear trend towards sO2 over-reading at the centre of the tube. This is a clear demonstration of spectral coloring. A second illustration of spectral coloring is shown in Fig. 8b where experiments were repeated with absorbing nigrosin dye incoporated into the background of the tissue mimicking phantom surrounding the flow circuit tube. Increasing the background absorption causes increasing inaccuracy of the sO2 relative to the ground truth since wavelengths of light are preferentially absorbed by the nigrosin dye, leading to over-estimation of the contribution of oxyhemoglobin to the PA signal. A simple light fluence correction, implemented by dividing the images by the known nigrosin spectrum (section 2.4), restored consistent sO2 measurements irrespective of the background absorption, providing evidence that this spatial phenomenon is indeed related to spectral coloring.

4 Discussion

The application of PAT in assessment of blood hemoglobin concentration and oxygen saturation is now widespread, but estimation of the oxy- and deoxy-hemoglobin chromophore concentrations is affected by a variety of tissue- and instrumentation-dependent factors. We have developed and applied a low-cost flow phantom system with online monitoring of optical absorption spectra and partial pressure of oxygen in order to facilitate testing of PAT systems.

We first selected an appropriate tubing to maximise PAT image quality and tested the circuit using two optically absorbing dyes. For both dyes, MB and ICG, the absorption spectrum exhibits a double peak where the relative peak intensities shift with increasing concentration. Unmixing with a single “average” spectrum is therefore unlikely to yield the same result as unmixing with a concentration-specific (online/offline) spectrum; indeed, we observed different and slightly non-linear relationships between dye concentration and spectrally unmixed intensities (which are assumed to be proportional to concentration) for the two types of unmixing. For ICG, the values from unmixing with the literature spectrum under-read those from unmixing with the online/offline spectra (Fig. 5b,d); for MB the unmixing results were indistinguishable (Fig. 5a,c) since the concentration dependent effects occur outside of the wavelength range of the PA system under test. It is possible that erroneous concentration measurements may also be caused by spectral coloring, even though only a single component is used in the spectral unmixing; two component unmixing (Hb and HbO2) is more relevant for calculation of sO2.

We proceeded to evaluate the estimation of sO2 under both closed circuit and dynamic conditions. The results highlight two major factors that contribute to inaccuracies in spectral unmixing, and therefore in sO2 measurements. The first is the choice of spectra for unmixing the photoacoustic images; the second is the effect of spectral coloring. The online spectrometer allowed spectra to be recorded during PAT image acquisition and used as alternatives to published spectra for unmixing the images. Moreover, the ground truth sO2 values provided by the spectrometer and the pO2 probes are designed to facilitate exploration of phenomena such as spectral coloring.

Images of blood flowing through the circuit entailed unmixing with two components: oxy- and deoxy- hemoglobin. The resulting sO2 calculations (HbO2/THb) were consistent over time but were dependent on the hemoglobin spectra used for unmixing, the blood oxygenation saturation, and also the position within the tube. The spectra for oxy- and deoxy- hemoglobin measured online were considerably different from those available in the literature that are widely used for spectral unmixing both in vitro and in vivo. Using our online flow spectrometer spectra improved the accuracy of the sO2 calculations compared to those made using the default spectra, suggesting that the online spectra were more representative of the actual optical absorption presented to the PAT system.

Over-reading of the lowest sO2 values was also observed and attributed to spectral coloring: the high absorption by deoxygenated blood at wavelengths below about $750\text{\,}\mathrm{nm}$ results in low light intensities at these wavelengths in the centre of the tube and therefore the PA intensities (proportional to both absorption and light fluence) varied with wavelength in a way more closely matching with oxyhemoglobin, leading to overestimation of the oxyhemoglobin concentration. Admittedly, these sO2 values are well below the physiological range ($70\text{\,}\mathrm{\char 37\relax}100\text{\,}\mathrm{\char 37\relax}$), nonetheless it is important to explore the full dynamic range in order to establish the accuracy limits of PA sO2 measurements. A simple example of correcting the effects of spectal coloring introduced by the tissue mimicking phantom surrounding the tube was achieved through dividing by the background nigrosin spectrum, which restored accurate sO2 measurements.

One limitation of the study is the discrepancy between our measured hemoglobin spectra and those widely used in the literature. This could be attributed to optical scattering due to the blood cells within the measurement cuvette, as literature studies normally measure the haemoglobin molecule directly; another factor could be the strong attenuation of the blood, necessitating use of a neutral density filter in order to prevent signal saturation when recording the reference spectrum (taken with PBS in the tube), but then removal of the filter for the blood measurements. Although the filter was taken into account in the calculation of abosrption spectra, the effect of blood scattering is difficult to compensate for; future experiments should therefore calibrate the measured spectra by extracting blood samples from the circuit at different sO2 values and validating the sO2 using a blood gas analyser. It would also be interesting to investigate oxygenation states of haemoglobin extracted from blood cells thereby obviating optical scattering effects. A further limitation arises from the pO2 measurements, as the dynamics of the deoxygenation study did not directly mirror the sO2 values obtained in the spectrometer. This may be due to an unrepresentative conversion from pO2 to sO2: the Severinghaus equation is derived from human, not mouse, blood data and also assumes certain values for parameters such as pH and temperature. Alternative conversions such as the Kelman equation were also explored, and measured values for pH and temperature were incorporated, but this did not significantly alter the trend in sO2 over time. It is possible that bubbles or insufficient contact with the blood in the circuit corrupted the pO2 readings, and therefore future versions of the circuit will integrate truly in-flow pO2 probes and a membrane oxygenator to avoid bubbles.

The application of the presented phantom opens a range of opportunities for future studies of tissue- and instrument-dependent correction factors in PAT. In particular, future work will attempt to model and account for the spectral distortion across the tube diameter observed during blood oxygenation studies. The development of corrections incorporating multiple choromphores inside and outside the tube both ex vivo and in vivo is more complex and continues to be investigated, for example using Monte Carlo simulations and model-based iterative minimisation [16, 17, 21, 22, 23, 19]. Moreover, the number and choice of wavelengths used for spectral unmixing is an important consideration [13] to be investigated in future work.

5 Conclusion

In summary, we have developed a low-cost flow phantom that includes an online spectrometer and partial pressure of oxygen probe to facilitate detailed validation of PAT measurements of blood oxygen saturation, sO2. We found that it is important to correctly identify the absorption spectra to be used for unmixing photoacoustic images in order to accurately determine even relative absorber concentrations. We also found that the co-dependence of photoacoustic signal intensity on light fluence and absorption leads to a major challenge in accounting for spectral coloring, which can lead to substantial underestimation of sO2. Further calibration and automation of the circuit will enable additional ex vivo studies requiring careful control and knowledge of blood sO2, as well as opening the possibility of exploring additional blood parameters such as pH, and also absorptive and other properties of different fluids. Detailed understanding of the photoacoustic signal origins ex vivo remains essential for proper interpretation of photoacoustic measurements made in vivo.

Disclosures

SEB has received research support from iThera Medical GmbH and PreXion Inc., vendors of photoacoustic imaging instruments.

Acknowledgements.

The authors would like to thank Ayaka Shinozaki for her assistance with data collection, James Joseph for his contributions to the experimental design, and Michael Schneider for his helpful input regarding data interpretation. This work was supported by Cancer Research UK (C47594/A16267, C14303/A17197) and the EPSRC-CRUK Cancer Imaging Centre in Cambridge and Manchester (C197/A16465 and C8742/A18097). We would like to thank the CRUK CI Core Facilities for their support of this work, in particular the Imaging Core and the Biological Resource Unit.