Detection of minimally scattered photons
It was recognised early in the history of optical imaging that photons which have been scattered a small number of times carry more spatial information than diffusive photons. Furthermore, if measurements could be made of minimally scattered photons, images could be reconstructed using the Radon transform as in x-ray CT, avoiding most of the difficulties associated with diffuse optical image reconstruction. Methods which can isolate minimally scattered photons from the diffusely scattered background, such as collimated detection, coherent techniques, and time-gating were reviewed in detail by Hebden et al. (1997). However, the fraction of minimally scattered photons transmitted across large (> several cm) thicknesses of tissue is immeasurably small, making this approach unsuitable for medical imaging. The length scale over which a collimated beam effectively becomes diffuse is known as the transport scattering length, which is about 1 mm in most biological tissues at NIR wavelengths. The transport scattering length is equal to the reciprocal of the transport scatter coefficient µ's, which is defined as , where µs is the scatter coefficient and g is the mean cosine of the scattering phase function.
Diffuse optical tomography can be contrasted with optical coherence tomography (OCT), which is a rapidly developing medical imaging modality, particularly in the area of ophthalmology. It is a range-gating technique which exploits the coherent properties of back-reflected light to generate very high spatial and temporal resolution images of tissues with a penetration depth of a few millimetres (Boppart 2003, Fercher et al. 2003, Fujimoto 2003).
Recently, the emphasis of research in medical imaging with diffuse light has moved away from the pursuit of high (~ mm) spatial resolution and towards functional imaging. It is widely appreciated that diffuse optical imaging can never compete in terms of spatial resolution with anatomical imaging techniques (e.g. x-radiography, ultrasound, and magnetic resonance imaging (MRI)), but offers several distinct advantages in terms of sensitivity to functional changes, safety, cost, and use at the bedside. In the following sections we review the alternative technological approaches currently being pursued, and discuss their relative merits and disadvantages.
Continuous wave systems
Measurements of the intensity of light transmitted between two points on the surface of tissue are not only relatively straightforward and inexpensive to obtain, but also contain a remarkable amount of useful information, as demonstrated by the clinical successes of NIRS (Obrig and Villringer 2003). So-called continuous wave (CW) systems require a source that either emits at a constant intensity, or is modulated at a low (a few kHz) frequency in order to exploit the significant improvements in sensitivity available from phase-locked detection techniques. CW sources have been used to investigate the head, testes and breast by viewing light which has been transmitted through the body since at least as far back as the early nineteenth century (Bright 1831, Curling 1843, Cutler 1929). CW transillumination of the breast (or “diaphanography”) received a brief revival of interest during the 1970s and 1980s with the introduction of NIR sources and detectors, but no significant clinical utility was demonstrated (Hebden and Delpy 1997)
Probably the most highly developed application of CW imaging technology is the study of haemodynamic and oxygenation changes in superficial tissues, and of the outer (cortical) regions of the brain in particular using optical topography. This typically involves acquiring multiple measurements of diffuse reflectance at small source-detector separations over a large area of tissue simultaneously or in rapid succession (see section 4.1). By keeping the separation small, measured signals are relatively high and therefore may be acquired quickly, enabling haemodynamic changes with characteristic responses as fast as a few tens of milliseconds to be studied. Thus, optical topography of the cortex represents a mapping technique analogous to electro-encephalography (EEG), which is sensitive to electrical activity in the cortex.
Recent technological advances have led to the development of arrays of individual sources and detectors which can be coupled to the head using flexible pads held in direct contact with the scalp. The availability of low cost, high power (several mW) and narrow bandwidth (< few nm) laser diodes over a broad range of NIR wavelengths have made them the popular choice of source for most optical imaging applications. The detector selected for a given application will depend on issues such as the desired sensitivity, stability, and dynamic range, as well as more practical concerns such as size and cost. A variety of semiconductor photodiodes are available, usually offering very good dynamic range at low cost. Avalanche photodiodes (APDs) are generally the most sensitive of the semiconductor detectors. However, for optimum sensitivity, photomultipliers (PMTs) are required, which can provide single- photon counting performance, although with a more limited dynamic range and at significantly greater cost. A thorough review of detection methods is given by Knoll (1999).
The number of distinct sources or detectors required can be reduced by switching components sequentially to different optical fibres within the array (known as multiplexing), at the expense of reduced temporal resolution. Most current systems use multiple detectors to record signals continuously in parallel, while sources are either activated sequentially, or are intensity-modulated at different frequencies simultaneously. In the latter case, detected signals from specific sources are isolated either by using lock-in amplifiers (Yamashita et al. 1999) or by Fourier transformation and appropriate filtration (Franceschini et al. 2003, Everdell et al. 2004). The former technique is used by the first commercial optical topography system, the Hitachi ETG-100 system. It consists of eight laser diodes at 780 nm and eight laser diodes at 830 nm, each modulated at a different frequency between 1 and 8.7 kHz (Yamashita et al. 1999). The sources are coupled in pairs to eight distinct positions on the subject, while light is detected at eight further positions using APDs. An array of 48 lock-in amplifiers is used to sample 24 distinct source-detector combinations at two wavelengths, and thus Hitachi refer to this device as a 24-channel system. The system has been widely evaluated for brain imaging (see section 4.1) where the sources and detectors are distributed over one or two lobes of the brain (Yamashita et al. 1999). A new model, the ETG-7000, is a 120-channel device which can image the entire adult cortex with 40 pairs of laser diodes and 40 APD detectors.
CW measurements have also been used for the considerably more challenging approach to imaging known as optical tomography, which involves generating a transverse slice or three-dimensional (3D) image (see section 3). Adequate sensitivity to deep tissues requires measurements at large source-detector separations, and consequently transmitted light must be integrated over periods of several seconds per source in order to obtain adequate signal. While this largely prohibits the analysis of short, isolated haemodynamic events, Schmitz et al. (2002) have demonstrated a CW tomography system which uses gating and averaging of the detected signals to reveal cyclic haemodynamic changes (Barbour et al. 2004). Their DYNOT (DYnamic Near Infrared Optical Tomography) system is currently marketed commercially by NIRx Medical Technologies (USA).
During the late 1990s, Philips Research Laboratories (Netherlands) began evaluating a breast tomography system based on CW measurements and a simple back-projection algorithm (Colak et al. 1999). The patient lay on a bed with her breast suspended within a conical chamber filled with a tissue-like scattering liquid. This approach had the very attractive benefit of eliminating the variability in surface coupling. The clinical performance of the Philips system was lower than desired, partly because of the inability of CW imaging to distinguish between internal absorbing and scattering properties (Arridge and Lionheart 1998). Similar commercial systems based on CW measurements have been developed by Imaging Diagnostic Systems Inc. (Grable et al. 2004), see Figure 1, and Advanced Research Technologies Inc. (Hawrysz and Sevick-Muraca 2000).
There are a number of disadvantages associated with CW imaging using absolute measurements of intensity:
- Intensity measurements are far more sensitive to the optical properties of tissues at or immediately below the surface than to the properties of localised regions deeper within the tissue. This is due to the characteristic “banana” shape of the volume of tissue over which the measurement is sensitive (known as the photon measurement density function or PMDF), which is narrow near the source and detector and very broad in the middle (Arridge 1995, Arridge and Schweiger 1995).
- The detected intensity is highly dependent on surface coupling. For example, an optical fibre moved slightly or pressed more or less firmly against the skin can result in a very large change in the measurement. The presence of hair or local variation in skin colour can also have a major influence on intensity measurements. Although means of calibrating for variable surface coupling have been implemented (see section 3.4.6), NIRS and imaging using CW sources have largely focussed on recording differences in intensity, acquired over a period short enough so that the unknown coupling can be assumed to have remained constant.
- Measurements of intensity alone at a single wavelength are unable to distinguish between the effects of absorption and scatter (Arridge and Lionheart 1998).
Alternative types of measurement have been explored in order to circumvent some of these problems inherent in CW data. The techniques which have demonstrated the most promise for imaging through larger thicknesses of tissue are those based on the temporal measurement of transmitted radiation, or an equivalent measurement in the frequency domain. These are now reviewed separately in the following two sections.
The temporal distribution of photons produced when a short duration (a few picoseconds) pulse of light is transmitted through a highly scattering medium is known as the temporal point spread function (TPSF). After travelling through several centimetres of soft tissue, the TPSF will extend over several nanoseconds. Early technical validation studies, performed using laboratory picosecond lasers and a variety of sophisticated and expensive detector systems, such as streak cameras, were discussed in the 1997 review (Hebden et al. 1997). Since then, advances in time-correlated single- photon counting (TCSPC) hardware and pulsed laser diodes have significantly reduced the cost and complexity of time-resolved measurement, and facilitated the development of multi-detector systems. TCSPC involves correlating the arrival time of a detected photon with the sampling of a known variable analogue signal. The difference between samples resulting from a detected photon and that from an external reference (derived directly from the source) provides a measurement of the photon flight time. An example of this technology is the time-to-amplitude converter (TAC). The technique generally requires a photon-counting photomultiplier tube (PMT) detector. Maximising temporal resolution requires a PMT with a minimum transient time spread (TTS) of photoelectrons across the tube. PMTs with a TTS of 150-200 ps are available, but for a significantly shorter TTS (< 50 ps) it is necessary to employ a microchannel-plate PMT (for examples, see www.hamamatsu.com). To date, the application of time-resolved measurement to diffuse optical imaging has involved two distinct approaches:
- A transillumination technique in which sources and detectors are arranged on 0pposite sides of a slab of tissue. Typically a single source and detector, aligned along a common axis, are scanned in two dimensions across each surface, and a single projection image is produced directly. The approach was used in the first demonstration of 2D time-resolved imaging of tissue-like media by Hebden et al. (1991), and the slab imaging geometry has been adopted for several breast imaging systems (Ntziachristos et al. 1998, Grosenick et al. 1999, Pifferi et al. 2003), described in more detail in section 4.3. An array of sources or detectors is often used so that off-axis measurements are also available, which provides a degree of depth information sufficient for a 3D image reconstruction (Ntziachristos et al. 1998, Grosenick et al. 2004b).
- A tomographic approach to imaging, which involves placing sources and detectors over the available surface of the tissue in order to sample multiple lines-of-sight across the entire volume either simultaneously or successively. Images are then reconstructed using techniques such as those described in section 3.
Early applications of time-domain measurements used time-gating to identify those photons which first emerge from the tissue, which are assumed to have travelled the shortest distance and therefore be least scattered. This approach is, however, limited by the number of available photons with sufficiently short flight-times. Experiments by Hebden and Delpy (1994) indicated that a degree of high resolution information is encoded into the shape of the TPSF, which can be extracted if the TPSF is measured sufficiently accurately. Time-gating techniques have also been developed which discriminate between late arrival photons which are predominantly affected by absorption, and early arrival photons which depend on both absorption and scatter (Grosenick et al. 2003). More recently, Selb et al. (2004) showed that time-gating can be used to provide additional depth resolution compared with CW measurements alone by rejecting light from superficial tissues.
Time-resolved measurements were first applied to clinical optical tomography by researchers at Stanford University (Benaron et al. 2000, Hintz et al. 2001), who developed an imaging system which was used to measure photon flight times between points on a newborn infant’s head (see section 4.2). However, since only a single solid state detector was used, transmitted light between each combination of source and detector position was recorded sequentially, resulting in scan times of between two and six hours. Much faster scan times are achievable by using multiple detectors, as demonstrated by the 32-channel time-resolved system developed at University College London (UCL) (Schmidt et al. 2000), and a 64-channel time-resolved imaging system built by Shimadzu Corporation in Japan (Eda et al. 1999).
The UCL system is based on TCSPC technology and a dual-wavelength fibre laser. The laser provides interlaced trains of picosecond pulses at 780 nm and 815 nm which are coupled to the surface of the subject via a 32-way optical fibre switch. Transmitted light is collected simultaneously by 32 detector fibre bundles, which deliver the light to four 8-anode microchannel-plate PMTs via 32 variable optical attenuators, which ensure that the intensity of detected light does not exceed the maximum photon counting rate of around 2.5 x 106 photons per second. The arrival time of each detected photon is measured with respect to a laser-generated reference signal, and TPSFs are accumulated.
A time-domain measurement can be equivalently expressed in the frequency domain. Some researchers have developed imaging systems which acquire frequency-domain data directly using a source that is amplitude modulated at a high frequency (a few hundred MHz), and measuring the reduction in amplitude and phase shift of the transmitted signal. While time-resolved measurements have the advantage of acquiring information at all frequencies simultaneously, frequency-domain systems can employ light sources and detectors which are significantly less expensive than those required for time-resolved systems.
The science and technology of phase measurement for NIRS and imaging has been comprehensively reviewed by Chance et al. (1998b). Systems require a radio- frequency (RF, typically a few hundred MHz) oscillator to drive a suitable laser diode and to provide a reference signal for the phase measurement device which receives the detected signal from an appropriate high-bandwidth detector (e.g. PMT or APD, depending on the desired sensitivity). Heterodyning is commonly performed to convert the RF to a few kHz prior to phase detection. The detected signal is digitized over an appropriate period of time, and phase and amplitude are computed.
The transillumination and tomographic approaches described in section 2.3 for time- resolved systems are equally applicable for frequency-domain devices, and both have been widely explored. In the mid-1990s two major companies in Germany reported development of breast imaging systems based on frequency-domain measurement of transmitted light. These prototypes, constructed by Carl Zeiss (Kaschke et al. 1994, Moesta et al. 1996) and by Siemens (Götz et al. 1998), both involved rectilinear scanning of a single source-detector pair over opposite surfaces of a compressed breast, resulting in single-projection images at multiple NIR wavelengths. Unfortunately the performance of both systems during quite extensive trials fell below that required of a method of screening for breast cancer, although various improvements to the Carl Zeiss system and their reconstruction method have since been implemented (Franceschini et al. 1997, Fantini et al. 1998).
As summarised in section 4, a variety of frequency-domain systems have been developed for both optical topography (Danen et al. 1998, Franceschini et al. 2000) and optical tomography (Pogue et al. 2001). Culver et al. (2003a) have built a hybrid CW/frequency-domain device for optical tomography which combines the benefits of speed and low cost of CW measurements with the ability to separate scatter and absorption available from the amplitude and phase of frequency-domain data. The system employs four amplitude-modulated laser diodes operating at different wavelengths (690 nm, 750 nm, 786 nm, and 830 nm) which are rapidly switched between 45 optical fibres on a 9 x 5 array. The array is positioned against one side of a compressed breast, while light emerging on the opposite side is focussed on to a CCD camera. Meanwhile, diffusely reflected light is detected simultaneously by APDs via nine fibres interlaced among the source fibres. The amplitude and phase of the APD signals are determined using a homodyne technique.
Comparison of time-domain and frequency-domain systems
The relative advantages and disadvantages of frequency-domain optical imaging systems compared to time-domain systems have been subject to debate for more than ten years. Frequency-domain systems are relatively inexpensive, easy to develop and use, and can provide very fast temporal sampling (up to 50 Hz). Time-domain systems, on the other hand, tend to use photon-counting detectors which are slow but highly sensitive. Hence, frequency-domain systems are well suited to acquiring measurements quickly and at relatively high detected intensities (such as for topography applications). However, when imaging across large (> 6 cm) thicknesses the light intensity is very low, possibly only a few photons per second, and can only be detected using the powerful pulsed laser sources and photon counting techniques incorporated into time-resolved systems.
The information content in a TPSF is inevitably greater than that in a single phase and amplitude measurement at one source frequency, but the magnitude of this benefit has yet to be thoroughly explored. The frequency content of the TPSF extends to several GHz, and while in principle a frequency-domain system could be designed to acquire this information, it is not yet possible to modulate high-intensity sources at such high frequencies. The intensity and mean photon flight time calculated from the TPSF are almost equivalent to the amplitude and phase of a frequency-domain system (Arridge et al. 1992). Other datatypes calculated from the TPSF such as variance, skew and the Laplace transform can provide enhanced separation between µa and µ's (Schweiger and Arridge 1999a), but may be more sensitive to noise. These additional datatypes have no simple equivalent in the frequency domain. Other approaches, such as time-gating to distinguish between µa and µ's (Grosenick et al. 2003) or to provide depth discrimination (Selb et al. 2004) also demonstrate the additional information available in the TPSF.
Practical optical measurements, particularly in the hospital, are commonly contaminated by constant (or temporally uncorrelated) background illumination. Frequency-domain systems are able to reject uncorrelated signals (but not the uncorrelated noise associated with these signals) by the use of lock-in amplifiers, while time-domain systems reject photons which reach the detector outside a finite temporal window. However, frequency-domain systems are unable to identify unwanted light which is temporally correlated with the measurement, such as light which has leaked around the object being imaged. In the time-domain, inspection of the TPSF can enable these contaminated measurements to be rejected (Hebden et al. 2004b).
Selection of optimal wavelength
The choice of wavelengths to use for NIR studies is a complex one. The “NIR window” used for tissue optics is bounded roughly between 650 nm and 850 nm. At lower wavelengths, absorption by haemoglobin limits penetration in tissue, while at higher wavelengths, absorption by water dominates. NIRS studies of the brain have typically employed wavelengths either side of the isobestic point of haemoglobin at 800 nm, where the specific extinction coefficients of HbO and HHb are equal. The actual wavelengths used for a given study are often determined in an ad hoc way and are often dictated by the availability of appropriate light sources (Cope 1991). Recently, Yamashita et al. (2001) Strangman et al. (2003), and Boas et al. (2004a) have shown experimentally and theoretically that a pair of wavelengths at 660-760 nm and 830 nm provides superior separation between HHb and HbO than the more commonly used 780 and 830 nm.
Pifferi et al. (2003) and Taroni et al. (2004a) have investigated the additional information obtained by using a wider range of wavelengths. They used four wavelengths (683 nm, 785 nm, 912 nm, and 975 nm) to image the breast, with the wavelengths selected empirically to optimise distinction between HbO, HHb, water and lipids. They also found that using four shorter wavelengths (637 nm, 656 nm, 683 nm, and 785 nm) improved the distinction between tumours and cysts.
The first systematic evaluation of the optimal wavelengths for NIR imaging was carried out by Corlu et al. (2003) who determined the wavelengths which minimised crosstalk between chromophores by maximising the uniqueness with which different chromophores can be distinguished. They simulated spatially varying distributions of HbO, HHb, water, and scattering coefficient, and determined the four wavelengths which minimised cross-talk between the chromophores. These did not correlate with the wavelengths which had been assumed to be optimal for NIRS studies. Such theoretical studies are of growing importance as advances in laser diode technology allow a much greater choice of wavelengths. This work also introduced the concept of reconstructing for chromophore concentration directly, rather than the more usual method of post-processing images generated at different wavelengths.
Uludag et al. (2004) used both theoretical and experimental methods to investigate the way in which cross-talk between calculated HHb and HbO concentrations ([HHb] and [HbO]) from dual wavelength measurements is affected by noise and error in the measurement. Their results agreed with those of Yamashita et al. (2001) and Strangman et al. (2003), i.e. that one of the selected wavelengths should be much shorter than 780 nm. They also showed that non-optimal wavelengths can lead to cross-talk which not only reduces the quantitative accuracy with which the changes can be determined but also changes the shape of the timecourse of the signal.
The predominant factor which reduces the image quality in diffuse optical imaging is scatter, and many of the advances in instrumentation, theory, and experimental techniques have been designed to reduce its effect. We will briefly mention two techniques which have been developed to improve the spatial resolution of diffuse optical imaging by combining the functional sensitivity of NIR measurements with the spatial resolution of ultrasound.
Ultrasound is an acoustic wave which can be used to modulate the amplitude of light travelling through a compressible object. This was exploited by Wang et al. (1995), who used a focussed ultrasound beam to modulate single channel optical measurements recorded across a cuvette. The amplitude of the modulated optical signal carried information about the optical properties at the focus of the ultrasound beam and was used to produce high-resolution images of the optical properties. Li et al. (2002) developed the technique further by imaging the output field with a CCD camera and analysing the laser speckle, which correlated with the ultrasonic modulation. They successfully imaged objects embedded in 25 mm thickness of tissue.
Photoacoustic imaging offers a further approach to combining the advantages of ultrasound and optical methods. When NIR light passes through tissue, it is absorbed, heating the tissue. The heated tissue expands and produces an acoustic wave which can be measured using conventional ultrasound detectors at the surface. An image of the sources of the ultrasound signal reveals the regions of highest optical absorption within the tissue, with the spatial resolution of ultrasound. Wang et al. (2003) have used photoacoustic techniques to produce anatomical and functional images of the intact rat brain with a spatial resolution of 200 ?m.