1. Introduction

Anomalous changes in peripheral blood flow are known to be a good indicator of various health disorders in the human organism. Laser Doppler imaging is a modern optical technique that allows accessing blood flow changes reliably and non-invasively [1]. The output of the instrument is a two-dimensional (2D) flow-related (perfusion, speed, concentration) map over an area of up to 50×50 cm². The technique is non-invasive because it involves no physical contact; the risk of infection and additional discomfort is completely avoided. However, present commercial laser Doppler imagers do not completely fulfill all requirements imposed on them by clinical applications. They are slow, special skills are required to use them, and the interpretation of the obtained results is not always objective. Thus, commercial laser Doppler imagers and the technique itself are mainly used in medical research projects but not often in clinical practice, despite the tremendous potential of laser Doppler imaging in the medical field.

Typical approach to improve the performance of these imagers, particularly to increase the imaging speed, is to parallelize measurements by using 1D or 2D arrays of photodetectors. Several blood-flow imaging systems have been designed over the last few years. Scanning laser Doppler imagers [2, 3], imagers based on speckle contrast analysis [4, 5] or laser speckle imagers (LSI) [6] were reported. A detailed review of these techniques can be found elsewhere [7]. However, there is still an important way to go for a widespread use in clinics, as existing instruments are either difficult to handle, slow, or not sufficiently accurate for measuring fast flow. Our goal was to develop a high-speed full-field laser Doppler imaging system that also should be reliable, objective, user- and patient-friendly.

To decrease the imaging time, a parallel detection scheme may be employed wherein the imaging speed increases by a factor proportional to the number of channels working in parallel. A 2D matrix of photodetectors is a suitable detection device for that purpose. Regarding photodetectors, there are presently four different technologies to produce photosensitive matrices: CCD, CID, simple PIN photodiode arrays, and finally CMOS. However not all of them can be employed for laser Doppler imaging. One limitation is imposed by the signal sampling frequency requirements. In laser Doppler blood flowmetry the measured frequencies are typically in the range of 0 to 20 kHz. The frame rate of CCD can hardly be faster than 1000 fps, which is not sufficient. Fast CCDs have been reported, however they are too much expensive for our application. CID technology allows for fast sub-frame rates however the sensitivity of these devices is better in the blue part of the spectrum, while laser Doppler blood flow measurements require red and near-infrared lasers. Besides the CID technology is very expensive. Use of an array of conventional PIN photodiodes is not attractive due to the insufficient packing density of the sensing elements that results in matrix being of a low resolution. However, the latest evolution of photodiode matrix technology, i.e. CMOS image sensors, possesses all the advantages for parallel detection of Doppler signals. These matrices have high resolution, good response in red, moderate response in near infrared, and they allow for fast sub-frame rates as pixels can be addressed randomly. Further, this technology is inexpensive and flexible as a result of the integration of photodetectors and associated electronics on the same chip.

In 2001 Serov et al. [8] have demonstrated the usefulness of the CMOS image sensor technology using a non-integrating CMOS sensor for measurements of blood flow by means of laser Doppler technique. Recently, a new generation of such an imager has been designed and applied to measurements of perfusion on human skin resulting in flow images every 90 seconds for a 256×256 pixel region of interest [9]. Both of these imaging systems were based on non-integrating CMOS image sensors. In this paper we describe a new laser Doppler imaging system that employs an integrating CMOS image sensor. The advantage of the integrating versus the non-integrating detector, particularly for laser Doppler imaging, is explained in the following section.

2. Integrating vs. non-integrating detectors

There exist two concepts in CMOS image sensor technology for recording the signals deposited by the photons in the detector: non-integrating and integrating. In non-integrating detectors the photon flux is continuously converted into an electrical output signal. To obtain images, the detector array is read instantaneously by means of sequential photoelectrical scanning. Each pixel detects only the photons that are received during the time the pixel is sampled:

Here T_tot is the time to read out all N pixels of the frame (or sub-frame). Thus, during ∆t one pixel detects X photons:

Here P_tot is the total optical power received by N photodetectors.

In the integrating detector concept the total deposited energy is integrated via charges accumulated while the detector is being hit by photons. The charges are accumulated in a small capacitor, which at the end of the time frame has to be read out. The charge is then converted into an output signal that is linearly proportional to number of photons that have hit the detecting pixel. Either each pixel collects photons during the time the other pixels are read out (rolling shutter mode), or all pixels collect photons during the integration time and they are read out immediately thereafter (global shutter mode). The maximum integration time is equal to the time to read N pixels, T_int =T_tot . Therefore, the number of photons detected by one pixel of an integrating detector array is

For both systems the signal to noise ratio (SNR) is determined by the number of detected photons X [10]:

So the net advantage in the SNR for the integrating image sensor is

Above, we compared two imaging systems, one with an integrating detector array and one with a non-integrating (scanning) detector array. We have assumed equal detector noise for both imagers, which is not always true. For completeness, the influence of the temporal noise on the SNR of each imaging system should also be considered.

For both types of sensors the minimum noise floor consists of thermal noise (TN) and shot noise (SN).

Here 〈I_photo 〉 is the average photocurrent and 〈I_dark 〉 is the average dark current in the circuit; k is the Boltzmann constant, T is the temperature in degrees Kelvin, B_n is the noise equivalent bandwidth, R is the load resistance, and q is the charge of an electron. The value of the load resistance is determined by the upper cutoff frequency required to pass the signal, f_s :

where C is the capacitance of the photodetector. The SNR is then

Here 〈$i_s^2$〉 is the mean-square value of the signal and M is the average number of speckles on a photodetector pixel [11].

In respect to the SNR, we first consider the class of non-integrating devices. In general, the noise bandwidth and the signal bandwidth are not the same. If the upper cutoff frequency is determined by a single RC time constant, then the signal bandwidth and the noise bandwidths are respectively

Thus for the non-integrating detector the SNR is

Second, for the integrating detector, the SNR is expressed as before (eq.(8a)) except that the noise bandwidth is now defined as B_n = 1/(2·T _int), where T_int is the time interval between successive readouts of the diodes (the integration time). Therefore, to match the signal bandwidth the integration time is determined by

Now we find that the SNR of the integrating detector is

Thus, at the same photocurrent, the SNR of the integrating detector is about a factor of 1.5 better than for the non-integrating device.

Finally, from eqs.(5) and (12) we find that, compared to the non-integrating detector case, where only one pixel of the image is measured at a time, the SNR of the integrating detector array can be increased by a factor of up to

The above considerations concern the fundamental difference between the detectors, however the technological features that influence the detector performance should also be mentioned. One problem encountered in non-integrating detectors is the dependence of the time constant on the signal level; this causes the non-integrating detector bandwidth to be dependent on the signal level. This problem could be eliminated in principle, but at the expense of an increased noise floor for the associated amplifier circuit [12]. As for the integrating system, an additional advantage available here is the possibility of reducing the effect of the thermal noise. This can be achieved by a well-known correlated double sampling signal processing method [13]. Also, the readout noise of the non-integrating sensor is usually about an order of magnitude higher than for the integrating one.

Another essential advantage of the integrating detector concept is the flexibility in selecting the integration time to always match the required signal bandwidth. Since both shot and thermal noises are distributed over a wide frequency range, reducing the noise bandwidth effectively reduces the noise in the measurement. Therefore the integration time can be used as an additional degree of freedom.

3. Experimental configuration; design considerations

Not any CMOS image sensor can be utilized for laser Doppler blood flow measurements. A limitation arises from the electronic architecture of a particular photosensor matrix. The main requirement here is a possibility to selectively read out the pixels from a predefined sub-frame at high-speed. Ideally, the sub-frames would be acquired at frame-rate of up to 40,000 frames per second as assumed for the maximum sampling frequency in laser Doppler flowmetry. Another important requirement is the spectral response of the sensor. For laser Doppler blood-flow measurements the source wavelength should be in the red to near-infrared range. Thus the spectral response of the detector should be optimized for this range. For our imager a digital CMOS camera based on the VCA1281 monochrome CMOS image sensor from Symagery (Canada) was utilized. This sensor operates in rolling shutter mode; it has a 1280H × 1024V resolution, a 7×7μm² pixel size, a 40 MHz sampling rate, and an 8-bit ADC. The sensor has a specified flat spectral response in the range between 500 and 750 nm. The camera was connected to the host PC via a fast LVDS (Low-Voltage Differential Signaling) interface providing for a high-speed transfer of the obtained frames.

For the sample illumination we used a solid-state-diode-pumped laser of 250 mW output optical power emitting at 671 nm. The laser beam was coupled to a ∅1.5mm plastic optical fiber. A GRIN (gradient index) lens of ∅1.8mm was placed at the distal end of the fiber. This configuration produced a uniform illumination of the sample. The illuminated area was up to ∅170mm. The intensity profile of the illuminating beam is shown in Fig. 1b (bottom). A slight increase of the intensity for higher pixel numbers is caused by the illuminating geometry – an angle of 9° between the illumination and observation directions. The high-frequency variations of the intensity are due to speckle effect.

The backscattered light was collected with an f=6 mm objective with an f-number of f#=1.2. The low f-number objective provided the system with the superior photon collection efficiency that becomes critical for short integration times (in the range of a few tens of milliseconds). The imager head was installed on the articulating arm system for providing an easy access to the object of interest, as shown in Fig. 1. Typically, the imager head was placed at a distance of 150-250 mm from the measured surface.

 

Fig. 1. a) High-speed laser Doppler imaging system. The imager head was mounted on an articulating arm to simplify access to the measured objects. b) Block diagram of the laser Doppler imaging system modules (top); and the intensity profile of the illuminating beam (bottom).

Download Full Size | PDF

In general, the geometry of the illuminating beam and the objective lens with respect to the sample influences the Doppler beat frequency response of the LDI system. For our system design we only found a minor influence (standard deviation less than 15%) for the imager head position within ±30° relatively to the vertical detection position. Typically, in optically dense biological tissues, such as skin, scattered photons loose their initial directions due to diffusion rapidly. Finally, the scattering angles as well as the directions of moving blood cells are random resulting in a less sensitive system response due to variations of the imaging geometry.

For the imager we have developed software with a simple and functional interface to control the system. The software allows for changing the sensor parameters, for control of the data acquisition mode, for acquisition of the data, and for display of the flow-related (perfusion, concentration, speed) maps. A photographic image of the sample and flow related maps displayed on the monitor are obtained with the same image sensor; therefore the obtained flow-maps can be easily associated with an area of interest on the sample. The flow-map data can be stored as separate images or as a movie. The imager can run in different modes: i) simple imaging mode, recording a single flow-map image, and ii) continuous imaging mode, with successive recording of flow-map images.

4. Data acquisition and signal processing

The signal sampling frequency is inversely proportional to the time to acquire one sub-frame. The sub-frame sampling rate of the sensor depends on its size and the pixel clock frequency. The clock frequency was fixed at 40 MHz for optimum performance speed/quality; the higher pixel-sampling rate increases the noise level. The size of the sampled sub-frame finally defines the signal sampling frequency of the imager. For 256×4 pixels sub-frame the frame sampling frequency was 30kHz, 256×6 pixels – 20 kHz, 256×8 pixels – 14 kHz, etc.

To obtain one flow map over a region of interest (ROI), which in our case was 256×256 pixels, the ROI must be subdivided in smaller regions (e.g. into 32 sub-frames of 256×8 pixels) and scanned electronically. From 32 to 512 sampled points were obtained for the acquired time-domain signal for each pixel of the sub-frame, thus the intensity fluctuation history was recorded for each pixels of the ROI.

The signal processing comprises the calculation of the zero- (M₀ ) and the first-moment (M₁ ) of the power density spectrum S(ν) of the intensity fluctuations I(t) for each pixel. The zero-moment is related to the average concentration, <C>, of moving particles in the sampling volume. The first moment (flux or perfusion) is proportional to the root-mean-square (rms) speed of moving particles, V_rms , times the average concentration [14]. The governing expressions are:

Here the variable ν is the frequency of the intensity fluctuations induced by the Doppler shifted photons. We calculated the power density spectrum using an FFT algorithm applied to recorded signal variations at each sampled pixel of the ROI. Noise subtraction is performed upon the calculated spectra by setting a threshold level on the amplitude of the spectral components. This filtering is applied to reduce the white noise (e.g. thermal and readout noises) contribution to the signal. Thereafter the perfusion, concentration and speed maps are calculated and displayed on computer monitor.

The signal processing is performed with a general purpose PC CPU – AMD-64-3000+. The total imaging time (including data acquisition, processing and display) depends on the number of samples obtained for each pixel and the ROI size. For 256×256 pixel ROI the imaging time is 2.5 sec for 64 samples, 3.5 sec for 128 samples, 5.5 sec for 256 samples, and 10 sec for 512 samples.

5. Measurements and results

In this section we present the results of the measurements obtained with our new imager. First, we made some in vitro measurements to characterize the performance of the imaging system in terms of bandwidth, and the linearity of the imager response to velocity and concentration changes. Second, we demonstrate the performance of our imager in vivo and present typical flow-maps obtained by measuring microcirculation in human skin.

5.1. In vitro

To measure the imaging system bandwidth and the response linearity to concentration and velocity changes, a signal from a sine-modulated light emitting diode (LED) was measured at one pixel. The LED was connected to an analog output of a synthesized function generator (Stanford Research System, Model DS345).

 

Fig. 2. a) The M₁/M₀ (velocity) imager response as a function of the measured signal frequency. b) The √M₀ (concentration) imager response as a function of the measured signal frequency. c) The √M₀ (concentration) imager response as a function of the measured signal amplitude. d) The SNR of the system as a function of the integration time for measurements on finger and forearm skin; error bars represent standard error for each measured SNR.

Download Full Size | PDF

The ratio between AC and DC components of the light output power could be set at required values changing the offset and the amplitude of the LED driver current. The frequency of the signal could also be changed. The number of points for FFT was 128. The signal sampling frequency of the imager (2f_s ) was around 12 kHz. This corresponds to the bandwidth from DC to 6000 Hz with 100 Hz frequency resolution. The integration time was 82 μs. Experimental results of these in vitro studies are shown in Fig. 2. Figure 2(a) shows the speed response of the imager M₁/M₀ as a function of the input signal frequency. The input signal of 10% modulation depth, AC_rms/DC, was measured for the frequency range from 100 to 6500 Hz. A linear dependence of the M₁/M₀ imager response is found up to the Nyquist frequency, which matches well to theoretical expectation. Effectively, the measured (M₁/M₀)f_s value should be equal to the signal frequency, which is clearly seen from the results. After approximately 6000 Hz, the decay in the imager response is observed due to the aliasing effect. It should be noted that the digital image sensors do not usually include antialiasing circuitry in their design; therefore the aliasing effect is virtually unavoidable in the imager. An antialiasing filter must be employed before the signal is digitized. It is ineffective to apply a low pass filter on the digitized signal to eliminate aliasing because the effect occurs prior to the sampling process. Any aliasing effects would already be stored in the digitized signal and could not be removed by low pass filtering as the effects typically appear as low frequencies in the signal.

In Fig. 2(b) the AC_RMS/DC response of the imager as a function of the input signal frequency is shown. The AC_RMS/DC value is proportional to the square of the M₀ value. The decay in the √M₀ imager response is due to the non-zero integration time of the detectors. This dependence is very similar to the frequency response of a basic low pass filter RC-circuit with a time constant defined by eq.(9) (see also eq.(11)). A decay of a factor of 0.5 for an RC-circuit is typical at the high-frequency cut-off. For integrating sensors, the measured signal response near the cut-off frequency is even smaller being approximately of 0.7 of its maximum, see eqs.(9).

In Fig. 2(c) the imaging system √M₀ response to the amplitude changes of the input signal is shown. The input signal frequency was fixed at 3000 Hz. The imager signal amplitude response shows expected linear dependence. At low amplitudes of the input signal the imager response demonstrates a nonlinearity caused by noise. The results shown in Fig. 2(d) are discussed in the next section.

5.2. In vivo

In this section we demonstrate the performance of the imager in vivo. In Fig. 3, flow-related maps obtained on finger skin of a healthy person are shown. The images were obtained for the imager settings for the bandwidth from DC to 6000 Hz with 100 Hz resolution; the integration time is 82 μs. The total imaging time is around 3.5 seconds. A smoothing filter is applied to the row images: the value of each pixel shown was obtained by averaging the row-values of 8 neighboring pixels. The flow maps (perfusion, concentration, speed) are false-coded with 9 colors. This coding is relative and does not mean that the measured perfusion value coded by, e.g. red, is equal to the value coded by the red color for concentration or speed. The images clearly show the difference in speed and concentration distributions measured on the fingers. The lower value for the concentration signal measured on the nail is caused by the higher amount of non-Doppler-shifted photons reemitted from the relatively thick statically scattering nail tissue compared to the thin statically scattering stratum corneum layer of the skin. The signal measured on the nail shows a higher speed of the moving blood cells in the under-nail tissue. However, we cannot say here definitely whether this is because of the blood speed was really higher under the nail, or because the measured values were obtained under two different conditions (e.g. homodyne detection for the skin and heterodyne detection for the nail) that could make the signal response be different despite the same speeds of the objects in both cases. This ambiguity is a common problem for all laser Doppler or laser speckle imagers and still needs to be investigated. The black and white photographic image of the object of interest, see Fig. 3(d), is obtained with the same CMOS camera. This image is useful for determining the anatomical boundaries associated with the perfusion regions presented in the blood flow maps.

 

Fig. 3. Flow-related maps obtained with the new imager on finger skin (ROI=256×256 pixels): a) perfusion map [Low=1500 a.u.; High=3000 a.u.]; b) blood concentration map [Low=150 a.u.; High=300 a.u.]; c) flow speed map [Low=500 a.u.; High=1500 a.u.]; d) image of the object. The imaging area is 5.5×5.5 cm². The imaging time is 3.5 seconds in total.

Download Full Size | PDF

Figure 4 shows the perfusion map images obtained during an artery occlusion experiment. The imager settings were the same as for the measurements described for Fig. 3 and 4 above. This example demonstrates the performance of the imager in the continuous imaging mode. The images were taken subsequently every 3.5 seconds, 3.5 s being the imaging time. The selected maps are shown in the matrix of 4×3 images to see the perfusion changes before, during and after the occlusion. As expected, there is a decrease of the perfusion signal during the occlusion. After the occlusion is released the local perfusion raised above the initial value; this effect is known as reactive hyperemia, shortly after which the blood flow returned to the initial state.

Finally, we demonstrate one of the possible applications of the imager. We measured perfusion changes in skin after a stimulant cream was applied. This cream contains 1% benzyl nicotinate. This is a pharmaceutical/cosmetical substance that is used in creams against arthritis pain and a small amount of this cream on the skin stimulates blood flow within a few minutes. The effect of the cream is clearly seen on the images shown in Fig. 5. The blood flow starts steadily increasing approximately 1.5 minutes after the cream was applied. The increased blood flow in the spot was also followed by reddening of the skin that was observed by eye. The effect is temporal lasting 1–2 hours.

 

Fig. 4. Artery occlusion experiment recording repeated perfusion images in real-time (ROI=256×256 pixels). Numbers show time (in seconds) when the images were obtained: 0–6 s, before occlusion; 16 s, occlusion on; 22–29 s, occlusion stopped blood flow; 35 s: occlusion is released; 38–45 s, post-occlusive hyperemia; 48–64 s, restored perfusion level. The imaging area is 5.5×5.5 cm². Low=1500 a.u.; High=3000 a.u.

Download Full Size | PDF

We measured the SNR of the instrument for measurements on finger and forearm skin. The SNR was measured and calculated according to SNR = ${\mathit{\text{SignalAC}}}_{\mathit{\text{rms}}}^2$ / ${\mathit{\text{NoiseAC}}}_{\mathit{\text{rms}}}^2$ as a function of different integration times of the sensor. The results are shown in Fig. 2(d). The noise signal was estimated from a M₀ flow-map histogram obtained by imaging a statically scattering white Teflon object. The signal values were found from the M₀ flow-map histogram obtained by imaging of finger and forearm skin. The SNR is increasing for longer integration times due to the decrease of the noise bandwidth, B_n , (the integration time could also affect the signal bandwidth). The SNR for finger skin is approximately 1.5–2 times higher compared for SNR obtained on forearm skin; finger skin perfusion is known to be higher than that of forearm skin.

 

Fig. 5. Perfusion images obtained with the high-speed laser Doppler imager (ROI=256×256 pixels). The imaging area is 5.5×5.5 cm². The effect of the stimulant cream (from Induchem AG, Switzerland) is the increased blood flow in the area where the cream was applied. The cream was applied on the skin of the inner side of the forearm. Images show the blood flow changes trough time: at 90, 97, 110, 124, 138, and 152 seconds after the cream was applied to the skin. The imaging time is c.a. 3.5 seconds per image. Low=500 a.u.; High=2500 a.u. (the red bar on the latest perfusion image is caused by an accident artifact during the sensor readout).

Download Full Size | PDF

6. Conclusion and outlook

In this paper we described the design and performance of a new high-speed laser Doppler imaging system based on an integrating CMOS image sensor. The use of a 2D matrix of integrating photodetectors results in an increased SNR of the system compared to the analogous imagers based on non-integrating detectors. The bandwidth of the imager with integrating detectors can be adjusted to the bandwidth of the signal thus increasing the SNR of the measurement. The use of 2D matrix of integrating photodetectors is particularly important for the parallel detection modality.

We tested the imager in vitro and in vivo. The imager demonstrated a reliable performance and fast imaging speed: e.g. for a 256×256 pixel ROI the imaging time was 3.5 seconds for 128 points taken for the FFT. This time is essentially smaller compared to the imaging time of the scanning imagers. Current commercial systems, mainly due to their system design based on sequential mechanical scanning, need more than 5 minutes to obtain a flow-image of the same resolution. Here, we measured the changes of blood flow over an extended area of tissue of 5.5×5.5 cm² virtually in real time. In principle, for this imager design, the imaging area can be up to 15×15cm² for the illuminating laser power of 250 mW. High-resolution real-time laser Doppler imaging would provide physicians with additional functional information about microcirculation.

We submit that the imaging time of future LDI systems will be less than one second approaching more and more the imaging time of the laser speckle imaging (LSI) system currently accepted as the fastest (video-rate), e.g. [6]. The LSI systems obtain flow-related information by measuring the contrast of the image speckles. Effectively, the contrast values measured by LSI are directly proportional to the normalized M₀ value that is measured by laser Doppler with integrating photodetectors. The images shown in Fig. 3 demonstrate a difference between perfusion (M₁ ) and concentration (M₀ ) maps, the former being of primary interest and the latter being of secondary interest. Thus the objectiveness of the LSI approach is not really evident in general case. We submit that laser Doppler imaging provides more objective information rather that the LSI method since with laser Doppler technique the concentration and speed signals can be measured independently. In LSI these two signals are always mixed, and thus it can be hard to attribute an exact cause for the changes in the contrast signal [11]. Generally, the LSI approach functions more as an indicator of flow rather than an instrument for quantitative measurements of physiological phenomena. However both systems have the potential to be used in combination, which may lead to even better overall performance for both technologies. Fortunately, the performance of both methods can now be compared objectively by measuring the same sample with a single integrating CMOS image sensor. This is our future plan.

Considering that the medical doctors would use the imager, our main effort was to develop a reliable, objective, accurate, user- and patient-friendly high-speed imaging system to measure blood flow in various biological tissues. From our present results, we anticipate additional patient comfort resulting from measuring times of less than 5 s. Although the training and experience will still be needed in order for users to interpret the results, the measurements themselves could be implemented easily due to a superior instrument design in combination with simple, interactive software.

The new imaging modality demonstrated the paves for an accurate, inexpensive, and easy to use diagnosis for a widespread dissemination by laboratories and hospitals. The wide range of applications is one of the major challenges for a future application of the imager. High-resolution high-speed laser Doppler perfusion imaging is an innovative technique for diagnosis and assessing the treatment of diseases such as atherosclerosis, psoriasis, diabetes, skin cancer, allergies, cardiovascular diseases, skin irritation. The new technique could also be applied for burn assessment, wound healing, and plastic surgery.