2.1.

Experimental System

Figure 1 shows a block diagram of system instrumentation. The entire system can be controlled via a single computer. A combined light delivery and ultrasound probe is mounted on a voice coil positioning stage (VCS-10-023-BS-01, $1.0\text{-}\mathrm{in.}$ travel, $2.3\mathrm{lb}$ continuous force, $6.9\mathrm{lb}$ peak force, purchased from H2W technologies, Inc., Valencia, California). This stage is driven by a programmed motor controller (Elmo Harmonica HAR 5/60, Elmo Motion Control Inc., Nashua, New Hampshire) to oscillate at fixed scanning speed. A digital output bit from the motor controller is set to high for a short duration at the end of each scan trajectory and sent to a digital input-output (DIO) card (NI PCI-6542, National Instruments, Inc). This end-of-scan (EOS) trigger is used to mark the beginning of an image frame and trigger the DIO card to generate a pulse sequence that is sent to the ultrasound pulser/receiver (5073PR, Panametrics, Waltham, Massachusetts) and/or laser. The ultrasound pulser/receiver is capable of a repetition rate up to $10\mathrm{kHz}$ , and can provide energy up to $16\mu \mathrm{J}$ with amplifier gain up to $39\mathrm{dB}$ .

Fig. 1

Block diagram of the combined PA and Doppler ultrasound system: MC, motor controller; DIO, digital input-output card; L, laser; T, ultrasound transducer; P/R, ultrasound pulser/receiver; DAC, digital acquisition card; EOS: end-of-scan.

Currently, our system can provide several imaging modes, such as ultrasound (US) $B$ -mode, PA mode and Doppler US mode. In PA mode, the DIO card is triggered by EOS output and then generates a pulse sequence to trigger the laser flashlamp and $Q$ -switch. It was found that $Q$ -switch triggering in addition to flashlamp (FL) triggering was important to reduce laser pulse timing jitter. A $Q$ -switched Nd:YAG laser (Ultra UL 421111, Big Sky Technologies, Bozeman, Montana) with a maximum pulse repetition rate of $20\mathrm{Hz}$ was used to obtain a $10\text{-}\mathrm{ns}$ pulse of $532\text{-}\mathrm{nm}$ wavelength. To provide a longer wavelength, a tunable laser system was also used in this study. A tunable optical parametric oscillator (Surelite OPO Plus, Continuum, Santa Clara, California) was pumped by a $Q$ -switched Nd:YAG laser (Surelite III, Continuum, Santa Clara, California) with repetition rate of $10\mathrm{Hz}$ . The tuning range is from $410\text{to}710\mathrm{nm}$ . PA and Doppler US modes are combined by using the DIO card to interleave laser and US triggers so that images can be coregistered.

Radio frequency (rf) data received by the US pulser/receiver is digitized by an eight-chanel PCI (peripheral component interconnect) data acquisition card (CS22G8, Gage Cobra, Gage Applied Systems, Inc.) with $12\text{-bit}$ dynamic range and sampling rates as high as 125 Msamples/s. Acquisition triggers are provided by the sync out signals from the laser $Q$ -switch and US pulser/receiver.

The aforementioned light delivery probe is based on a confocal dark-field laser illumination approach, with some similarities to the dark-field PAM system first described by Maslov ² As shown in Fig. 2 , an incident laser beam is reflected by a right angle prism which is mounted on an acrylic tube with refractive index of 1.46. The downward light is diverted by a polished reflective cone to a $25\text{-}\mathrm{deg}$ polished surface of the tube. Then the reflected light travels downward along the tube and reaches another polished surface, which is machined at a certain angle to deflect the light confocally around the transducer focal axis. A $25\text{-}\mathrm{MHz}$ single element ultrasound transducer (V324-SM, $12.7\mathrm{mm}$ focal length, Panametrics, Waltham, Massachusetts) is positioned inside the tube and can be adjusted vertically to match its focus point with the laser focus. In this design, light is focused at $\sim 10.5\mathrm{mm}$ below the bottom of the probe.

Fig. 2

Scheme of light delivery probe: R, reflective cone; T, US transducer.

2.2.

Flow Phantom

Experiments were performed on flow phantoms consisting of transparent vessels of 0.2- and $0.86\text{-}\mathrm{mm}$ inner diameters (I.D.) embedded in a tissue-mimicking background. Polymer tubing (Paradigm Optics, Vancouver, Washington), composed of polyethylene terephthalate glycol (PETG) was used in construction of blood vessels. The inner diameters are representative of small arterioles and small veins.⁸ The tissue-mimicking base was mixed with 10% cornstarch and 10% gelatin by weight, which offered an optical reduced scattering coefficient ${\mu }_s^{\prime }=9.2{\mathrm{cm}}^{-1}$ , similar to human tissues. The ultrasonic and mechanical properties are also within the range of typical biological tissues referred to previous work.⁹

Blood-mimicking fluid was pumped at mean flow velocities of 1 and $5\mathrm{mm}∕\mathrm{s}$ using a calibrated syringe pump (NE-300 Syringe Pump, New Era Pump Systems Inc., Eantagh, New York). A $1\text{-}\mathrm{mm}∕\mathrm{s}$ flow speed is thought to be typical in small vessels in the microcirculation. This is also a speed that begins to challenge our Doppler microultrasound system. We simply chose $5\mathrm{mm}∕\mathrm{s}$ as a speed significantly above $1\mathrm{mm}∕\mathrm{s}$ , and could be representative of flow in venules or arterioles.⁸ The blood-mimicking fluid was composed of 5% by mass cornstarch dissolved in water, which ensures the blood to tissue signal power is consistent with in vivo signal levels. To increase the buoyancy of the particles, the mixture was first heated to $100°\mathrm{C}$ for several minutes and then cooled to room temperature before using. This fluid was also optically dyed by crystal violet to provide two optical absorption coefficients. One is $250{\mathrm{cm}}^{-1}$ at $532\mathrm{nm}$ , the other is $3.3{\mathrm{cm}}^{-1}$ at $630\mathrm{nm}$ , both of which are similar to human blood. The distance between the vessel and surface of the tissue phantom is $2\mathrm{mm}$ .

2.3.

Signal Processing Strategies

PA and power Doppler US images are coregistered and superimposed on $B$ -mode structural US images. To generate power Doppler images, steps in the signal processing chain involve signal alignment within and between lines of sight, wall filtering, Doppler power calculation, and thresholding.

In contrast to systems in which an ultrasonic pulse is repeatedly directed to a discrete line of sight, a swept-scan mode is used in our system to continuously scan over the flow phantom. The transducer is positioned over the region of interest, transmits one pulse and is laterally translated by one step within the pulse repetition period. This method can shorten the data acquisition time, but may cause signal decorrelations within each line of sight due to motion of the transducer. Three correlation-based alignment algorithms have been proposed for compensation purpose: single reference line, incremental, and averaging.¹⁰ The first strategy selects a single pulse within a line of sight as reference and aligns other pulses by the relative shifts. The second calculates the relative shifts of consecutive pulses and aligns each pulse with the previous one. The third involves averaging the correlation over a set of reference lines. The nature of the swept-scan mode results in it being suitable for an incremental method. Within one line of sight, the relative shift of consecutive pulses is determined by finding the maximum of the cross-correlation estimate. In our case, interlaced PA signals may aggravate the discontinuities between lines of sight, thus we extend this incremental alignment method to better visualize blood vessel continuity. To perform alignment between lines of sight, the relative shift between the last pulse of one line of sight and the first pulse of the subsequent line of sight is first determined. Then the entire block of pulses within the subsequent line of sight is shifted by this determined amount.

Wall filters are necessary to reject strong clutter signals from surrounding tissues before Doppler power calculation. Typically an infinite impulse response (IIR) high-pass filter is used to remove slowly changing signals corresponding to tissue echoes. The limitation of such filters is that they have poor performance when Doppler frequencies generated by tissue motion exceed the cutoff frequencies. Thus, eigenfilters have been proposed to base the clutter filter on the statistics of the clutter signal so that the response can be adapted to the tissue movement. To address the fact that the clutter signal is a nonstationary random process, the discrete Karhunen-Loeve transform (DKLT) is used to characterize clutter signals.¹¹ It decomposes the autocorrelation matrix of received rf data into eigenvectors and eigenvalues, and then eigenvalues are sorted in decreasing amplitude. The eigenvalue spectrum is a generalization of the Fourier power spectrum, which represents the energy of the signal components. Thus, we can partition clutter signals, blood flow signals, and white noise by their eigenvectors according to this energy spectrum and eliminate the clutter noise by choosing appropriate filter order.

After wall filtering, the intensity of signals from moving blood is accumulated within each Doppler ensemble to generate power Doppler image. By a Doppler ensemble, we mean a sequence of $N$ pulses, the echoes from which are processed to form one power Doppler $A$ -scan line.

2.4.

Receiver Operating Characteristic Analysis

Receiver operating characteristic (ROC) analysis can be used as a means of quantifying the performance of both power Doppler and PA imaging modes for the task of detecting blood (in our case blood-mimicking fluid). For each pixel in our images, we make a decision as to whether “blood” is present or absent. One way of making such a decision is to threshold the power Doppler or PA image. In practice, values above the threshold are color-coded (and thus classified as blood present) and the color-coded image is overlaid on a gray-scale US B-scan for structural context. By setting the threshold very low, a large number of color pixels may appear throughout the power Doppler (or PA) image due to noise, clutter, or other artifacts. Thus, a low threshold will yield a large fraction of true positives, but also a large number of false positives. A high threshold value may decrease the number of false positives, but at the expense of true positives. The trade-off between sensitivity and specificity can be graphically represented by an ROC curve that plots the true-positive fraction (TPF) versus the false-positive fraction (FPF) as the threshold level is varied from minimum to maximum. While we are interested ultimately in the estimation task of quantifying fractional blood volume in a region (which estimate can be thought of as the sum of the true positives and false positives), we note that thresholding can greatly impact this estimate. Moreover, to compare the performance of power Doppler and PA imaging modes, different threshold levels could be used for each mode. By comparing ROC curves for each mode (for different vessel sizes and flow conditions), we are able to objectively compare respective detection performances for the entire spectrum of possible threshold values. The area under the ROC curves (AROC) is used as a summary figure of merit to evaluate the detection performance of the two imaging modes.

3.1.

Flow Phantom Experiments

To demonstrate the performance of our combined system, flow phantoms with different vessel sizes were imaged by cross-sectional scans. Acoustic coupling was achieved by immersing flow phantoms in a water bath. The transducer was positioned so that the focus was $2\mathrm{mm}$ below the surface of the flow phantoms. The following acquisition settings were used in all experiments: $4\text{-cycle}$ pulse length, $39\text{-}\mathrm{dB}$ receiver gain, and $16\text{-}\mu \mathrm{J}$ transmit energy. The motion controller was programmed to drive the voice coil scanning at a speed of $0.5\mathrm{mm}∕\mathrm{s}$ , which provides a distance of $25\mu \mathrm{m}$ between each line of sight. System specifications and experiment conditions are listed in Table 1 . Scan speed and Doppler ensemble size are inversely related because the transducer is swept continuously during acquisition. The pulse repetition frequency (PRF) was set to $200\mathrm{Hz}$ when the laser was triggered at $20\mathrm{Hz}$ (the PRF was set to $100\mathrm{Hz}$ when the laser was triggered at $10\mathrm{Hz}$ ), yielding a Doppler ensemble size of 10. When acquiring interleaved PA and Doppler signals, the trigger delay of the data acquisition card was set to $6\mu \mathrm{s}$ and the length of acquisition window was set to $20\mu \mathrm{s}$ , which enables a capture of $8.5\text{-}\mathrm{mm}$ axial depth in both PA and power Doppler modes. The lateral distance of one scan trajectory was $10\mathrm{mm}$ . Interlaced data were separated to form individual PA, $B$ -mode and power Doppler US images. Then PA and power Doppler images were coregistered and superimposed on the $B$ -mode US images with structural contents.

Table 1

System specifications and experiment conditions.

Coregistered power Doppler and PA images acquired from the flow phantom with a $0.86\text{-}\mathrm{mm}$ vessel diameter are shown in Figs. 3 and 4 . Appropriate regions of interest (ROIs) are selected so that the blood volume fractions can be displayed in the center of the ROI. The ROI dimensions include a depth of $2.5\mathrm{mm}$ from top to bottom and a total lateral distance of $2.5\mathrm{mm}$ .

Fig. 3

Coregistered power Doppler and PA images of the flow phantom with a $0.86\text{-}\mathrm{mm}$ vessel size: (a) power Doppler image at $5\mathrm{mm}∕\mathrm{s}$ , (b) power Doppler image at $1\mathrm{mm}∕\mathrm{s}$ , and (c) PA image at $532\mathrm{nm}$ .

Fig. 4

Coregistered power Doppler and PA images of the flow phantom with a $0.86\text{-}\mathrm{mm}$ vessel size: (a) power Doppler image at $5\mathrm{mm}∕\mathrm{s}$ , (b) power Doppler image at $1\mathrm{mm}∕\mathrm{s}$ , and (c) PA image at $630\mathrm{nm}$ .

Figure 3 shows power Doppler images acquired at mean flow velocities of 5 [Fig. 3] and $1\mathrm{mm}∕\mathrm{s}$ [Fig. 3], and coregistered PA images acquired at $532\mathrm{nm}$ [Fig. 3]. In power Doppler mode, the cross section of blood vessel is well imaged at a high flow velocity [ $5\mathrm{mm}∕\mathrm{s}$ , Fig. 3], but hard to interpret at $1\mathrm{mm}∕\mathrm{s}$ due to color pixel artifacts [Fig. 3]. In PA mode, only the top surface can be well visualized because of the minimal light penetration depth in the vessel [Fig. 3]. (Color online only.)

To show longer penetration depth in the PA image, we repeated our experiment on $0.86\text{-}\mathrm{mm}$ flow phantom using a $630\text{-}\mathrm{nm}$ wavelength. Figure 4 shows the PA image acquired at $630\mathrm{nm}$ [Fig. 4] and coregistered power Doppler images [Fig. 4 and Fig. 4]. In comparison to Fig. 3, Fig. 4 shows that a deeper penetration has been achieved, but the entire cross section of the vessel is still hard to see.

Figure 5 shows the coregistered power Doppler and PA images acquired from the flow phantom with a $0.2\text{-}\mathrm{mm}$ vessel diameter. The ROI is $1.2\times 1.0\mathrm{mm}$ . The mean flow velocities are 5 [Fig. 5] and $1\mathrm{mm}∕\mathrm{s}$ [Fig. 5]. The optical wavelength is $532\mathrm{nm}$ . As shown in Figs. 5 and 5, power Doppler imaging has a good detection performance at a small vessel size $\left(0.2\mathrm{mm}\right)$ and high flow velocity, but poor performance at slow flow velocity $(1\mathrm{mm}∕\mathrm{s})$ . However, this small vessel can be easily detected in the PA image [Fig. 5].

Fig. 5

Coregistered power Doppler and PA images of the flow phantom with a $0.2\text{-}\mathrm{mm}$ vessel size: (a) power Doppler image at $5\mathrm{mm}∕\mathrm{s}$ , (b) power Doppler image at $1\mathrm{mm}∕\mathrm{s}$ , and (c) PA image at $532\mathrm{nm}$ .

3.2.

Image Analysis

Figure 6 shows ROC curves generated from both power Doppler and PA images at different vessel sizes, mean flow velocities, and optical wavelengths. Inspection of these individual ROC curves provides objective evaluation of blood flow detection performance of our combined system. Figure 6 shows the ROC analysis of power Doppler and PA images acquired at $0.86\mathrm{mm}$ I.D. and $532\mathrm{nm}$ . The color thresholds are varied through the range of values, which ensures that the full ROC curve is populated. As shown in Fig. 6, the power Doppler image acquired at $5\mathrm{mm}∕\mathrm{s}$ produces an ROC curve that is consistently higher than the PA-ROC curve, indicating that power Doppler can provide a better estimation of blood volume fraction at $5\mathrm{mm}∕\mathrm{s}$ compared to the PA mode. The ROC curve generated from $1-\mathrm{mm}∕\mathrm{s}$ power Doppler image moves close to a $45\text{-}\mathrm{deg}$ diagonal line, which means that the detection accuracy and sensitivity decreases at low flow velocity. Figure 6 shows the ROC analysis of power Doppler and PA images acquired at $0.86\mathrm{mm}$ I.D. and $630\mathrm{nm}$ . The PA-ROC curve generated from $630\mathrm{nm}$ is higher than the one from $532\mathrm{nm}$ , which is due to the longer penetration depth. Figure 6 shows the ROC analysis when the vessel diameter is $0.2\mathrm{mm}$ . While the ROC curve generated from the $1\text{-}\mathrm{mm}∕\mathrm{s}$ power Doppler image lies almost along the diagonal line, the PA-ROC curve traverses through more of the upper left-hand quadrant of the ROC space. This is consistent with what we observe from the images, i.e., that PA imaging provides a better blood volume detection performance compared to power Doppler at small vessel size and slow flow velocity.

Fig. 6

ROC curves generated from PA and power Doppler images acquired at different vessel sizes and optical wavelengths: (a) $0.86\mathrm{mm}$ and $532\mathrm{nm}$ , (b) $0.86\mathrm{mm}$ and $630\mathrm{nm}$ , and (c) $0.2\mathrm{mm}$ and $532\mathrm{nm}$ . TPR, true positive rate; FPR, false positive rate.

The AROC values at different conditions are calculated using a trapezoidal rule and are shown in Table 2 . When the vessel diameter is $0.86\mathrm{mm}$ , the AROCs demonstrate that good estimation of blood volume fraction, defined by AROC $>0.90$ , can be achieved by power Doppler image at a high flow velocity $(5\mathrm{mm}∕\mathrm{s})$ . For a small vessel size of $0.2\mathrm{mm}$ and a slow flow velocity of $1\mathrm{mm}∕\mathrm{s}$ , the PA image has an AROC of 0.95, indicating a good performance of blood volume detection.

Table 2

AROC for PA and power Doppler images.