Three-dimensional optical tomography of the premature infant brain

INSTITUTE OF PHYSICS PUBLISHING
PHYSICS IN MEDICINE AND BIOLOGY
Phys. Med. Biol. 47 (2002) 4155–4166
PII: S0031-9155(02)39986-X
Three-dimensional optical tomography of the
premature infant brain
Jeremy C Hebden1, Adam Gibson1, Rozarina Md Yusof1,
Nick Everdell1, Elizabeth M C Hillman1,4, David T Delpy1,
Simon R Arridge2, Topun Austin3, Judith H Meek3 and John S Wyatt3
1 Department of Medical Physics & Bioengineering, University College London, 11-20 Capper
Street, London WC1E 6JA, UK
2 Department of Computer Science, University College London, Gower Street, London WC1E
6BT, UK
3 Department of Paediatrics and Child Health, University College London, 5 University Street,
London WC1E 6JJ, UK
Received 26 July 2002
Published 12 November 2002
Online at stacks.iop.org/PMB/47/4155
Abstract
For the ?rst time, three-dimensional images of the newborn infant brain have
been generated using measurements of transmitted light. A 32-channel time-
resolved imaging system was employed, and data were acquired using custom-
made helmets which couple source ?bres and detector bundles to the infant
head. Images have been reconstructed using measurements of mean ?ight time
relative to those acquired on a homogeneous reference phantom, and using a
head-shaped 3D ?nite-element-based forward model with an external boundary
constrained to match the measured positions of the sources and detectors.
Results are presented for a premature infant with a cerebral haemorrhage
predominantly located within the left ventricle.
Images representing the
distribution of absorption at 780 nm and 815 nm reveal an asymmetry consistent
with the haemorrhage, and corresponding maps of blood volume and fractional
oxygen saturation are generally within expected physiological values.
1. Introduction
Perinatal hypoxic-ischaemic brain injury is a major cause of permanent disability in very pre-
term infants who survive after neonatal intensive care. While new intervention methods are
being developed which could reduce the incidence of serious handicap, there are no reliable
methods of assessing either the degree of injury or the effectiveness of such procedures
which can be used safely and continuously on infants in intensive care. The need for an
effective method of non-invasively and quantitatively assessing regional cerebral oxygenation
4 Present address: Argose Inc., Waltham, MA 02451, USA.
0031-9155/02/234155+12$30.00
© 2002 IOP Publishing Ltd
Printed in the UK
4155

---

4156
J C Hebden et al
and perfusion in critically-ill infants has motivated researchers to explore methods based on the
transmission of visible and near-infrared light. The application of near-infrared spectroscopy
(NIRS) to continuously monitor cerebral oxygenation and haemodynamics non-invasively
was ?rst reported by J¨obsis (1977). This technique exploits the optical properties of natural
chromophores and of haemoglobin in particular, which has oxygenated and deoxygenated
forms with different characteristic absorption spectra in the near-infrared wavelength range.
NIRS was ?rst used to study the cerebral oxygenation of newborn infants by Brazy et al (1985),
who measured changes in both haemoglobin and cytochrome aa3 absorption which revealed a
strong correlation with oxygenation measured transcutaneously. Quantitative measurements
were ?rst reported by Reynolds and colleagues (Reynolds et al 1988, Edwards et al 1988,
Wyatt et al 1986) who derived various oxygenation and haemodynamic parameters on sick
newborn infants, including changes in (oxygenated, deoxygenated and total) haemoglobin
concentrations, cerebral blood volume and cerebral blood ?ow. The use of NIRS for studying
infant cerebral haemodynamics and metabolism has become quite widespread (e.g., Skov et al
1991, Wyatt 1993, du Plessis 1995, Meek et al 1999, Isobe et al 2000) and the technique has
also been employed to monitor the human foetal brain during labour (D’Antona et al 1997).
While NIRS has proven to be highly effective at monitoring tissue optical properties
at speci?c sites on the infant head, methods are being pursued to generate images via
measurement at multiple sites. The most straightforward imaging method is known as optical
topography, which involves acquiring multiple re?ectance measurements at small source–
detector separations over a large area of the head simultaneously or in rapid succession.
By keeping the separation low, measured signals are relatively high and therefore may be
acquired quickly, enabling fast haemodynamic changes to be studied. Optical topography of
the premature infant cortex was ?rst performed by Chance et al (1998), using a small array of
detectors and sources. Images were acquired from the head of a pre-term newborn infant which
revealed the response in the contralateral hemispheres to touching of the left and right hand.
A commercial optical topography system developed by Hitachi Medical Corporation (Tokyo,
Japan) has been employed to investigate spontaneous changes in the cerebral oxygenation
state of newborn infants during sleep (Taga et al 2000), revealing periods of oscillation in
concentrations of oxy- and deoxy-haemoglobin. Even more recently, Hintz et al (2001) have
mapped the cerebral cortex of the premature infant brain during passive motor activation.
An array of source and detector ?bres was used to reveal contralateral changes in cerebral
absorption at two wavelengths in response to speci?c movement of the infant arm. An iterative
image reconstruction procedure (Boas et al 2001a) was employed to estimate the distribution
of absorption within a thin slab of tissue at a speci?c depth below the probe, although an
assumption was necessary that all optical properties above and below the imaging plane
remain unchanged.
A signi?cantly more challenging imaging method which offers the facility to generate
images of the entire 3D volume is known as optical tomography. The sensitivity to deep tissues
requires measurements at large source–detector separations, and consequently transmitted light
must be integrated over periods of several seconds (or longer) per source in order to obtain
adequate signal. This largely prohibits the analysis of fast haemodynamic and metabolic
phenomena, but enables long-term oxygenation changes throughout the entire brain to be
monitored over a period of hours or days. Optical tomography is based on the general
principle that a ?nite set of measurements of transmitted light between pairs of points on the
surface of an object is suf?cient to reconstruct a transverse slice or 3D volume representing the
distribution of internal scatterers and absorbers. A variety of technologies have been explored
for acquiring suitable data for optical tomography. Some current optical tomography systems
are based on straightforward measurements of transmitted intensity (e.g., Schmitz et al 2001,

---

Three-dimensional optical tomography of the premature infant brain
4157
Siegel et al 1999). However, concerns that intensity data have an overwhelming dependency
on surface interactions have led many investigators to develop instruments which perform
measurements in the time (e.g., Eda et al 1999, Ntziachristos et al 1998) or frequency (e.g.,
Franceschini et al 1997, McBride et al 2001) domain. The former measure the temporal
distribution of photons transmitted between points on the surface in response to illumination
by an impulse of light, while the latter determine the modulation amplitude and phase delay
in response to an intensity-modulated signal. The ?rst research group to demonstrate optical
tomography of the neonatal brain was that at Stanford, who have constructed an imaging
system which measured photon ?ight times between points around the circumference of the
infant head (Benaron et al 2000, Hintz et al 1998). Scans have been performed on infants
at a variety of gestational ages, principally in order to establish the ability of the method to
identify intracranial haemorrhage (Hintz et al 1998, 1999). Two-dimensional (2D) images
are reconstructed using a simple backprojection method (Benaron et al 1994). A blinded
clinical comparison with established diagnostic methods (such as ultrasound) yielded a correct
interpretation of the optical images in six out of eight cases (Hintz et al 1999). A study has
also revealed focal regions of low oxygenation after acute stroke (Benaron et al 2000). The
ability of NIR measurements to identify intracranial haemorrhage has also been previously
explored by Robertson et al (1997), who performed a large number of spectroscopic studies
on adults and children with traumatic head injury. A high correlation with ?ndings on CT
scans was demonstrated, although their system was unable to detect deep intraparenchymal
haematomas.
The ground-breaking experiments by the Stanford group suffer from a number of serious
limitations, such as the use of a single detector and a low source power (100 µW average),
which required scan times of between 2 and 6 h (depending on head size and number of
wavelengths used). The low power also limits the maximum source–detector separation to
about 5 cm, and thus restricts the available information about optical properties near the
centre of the infant brain. Furthermore, the simplicity of the image reconstruction algorithm
employed by the Stanford group ignores the inherent 3D nature of photon migration in tissues,
and the highly heterogeneous structure of the human head. A more legitimate approach to
imaging complex tissues is to determine the parameters which describe an appropriate model of
photon transport within the investigated medium by comparing its predictions with measured
data (Arridge 1999). The model is then adjusted iteratively until acceptable correspondence
is achieved, and convergence towards the correct solution is assisted by the use of appropriate
regularization methods. This technique requires three distinct components: a forward model
which can generate a set of reliable measurements from a given 3D distribution of scattering
and absorbing parameters; the de?nition of an objective function to be minimized, based on
the error between model predictions and experimental data; and a scheme for adjusting the
parameters of the forward model to achieve the minimization. This method is the basis of the
algorithm known as TOAST (temporal optical absorption and scattering tomography) which
has been developed at UCL by Arridge and Schweiger (1997). TOAST employs a ?nite
element method (FEM) forward model, and uses an iterative model-?tting routine wherein
the FEM model parameters are repeatedly updated to optimize the match of the model to
the data. As described in the following section, a multi-channel time-resolved system has
been constructed at UCL speci?cally for the task of imaging the newborn infant brain and
to exploit the TOAST reconstruction package. Dual-wavelength measurements of the time-
resolved transmittance between multiple pairs of points on the surface of the head are used to
reconstruct images sensitive to local variation in chromophore absorption and tissue transport
scattering coef?cient. In this paper we report our ?rst 3D optical tomography images acquired
from a newborn infant in intensive care.

---

4158
J C Hebden et al
2. Instrumentation
2.1. MONSTIR
The UCL imaging system, known as MONSTIR (multi-channel opto-electronic near-infrared
system for time-resolved image reconstruction), is based on time-correlated single photon
counting (TCSPC) instrumentation, and is described in detail by Schmidt et al (2000a). A
custom-built light source was recently incorporated which consists of two ?bre lasers operating
at 780 nm and 815 nm. This provides interlaced trains of picosecond pulses at a 40 MHz
repetition rate (per pair of pulses) and a mean power of up to 55 mW per wavelength. The
pulses are coupled to the surface of the patient via a 32-way optical ?bre switch. Transmitted
light is collected simultaneously by 32 detector ?bre bundles, which deliver the light to four
eight-anode microchannel-plate photomultiplier tubes (MCP-PMTs) via 32 variable optical
attenuators (VOAs), which ensure that the intensity of detected light does not saturate or
damage the MCP-PMTs. The arrival time of each detected photon is measured with respect
to a laser-generated reference signal, and histograms of photon ?ight times (temporal point
spread functions, or TPSFs) are accumulated. The full set of TPSFs is subsequently transferred
to a dedicated workstation for processing and image reconstruction. In combination with the
TOAST reconstruction software, MONSTIR has been used to image a variety of tissue-
equivalent phantoms (Hebden et al 1999, 2001, Schmidt et al 2000a, 2000b), the human
forearm (Hillman et al 2001) and the female breast (Hebden et al 2002a).
2.2. The ?bre holder helmet
Acquiring measurements of light transmitted between multiple locations on the surface of a
newborn infant head represents a signi?cant challenge, particularly in the case of premature
and/or critically-ill infants. The design of the interface must take into account the comfort
and safety of the infant, and must not interfere with normal treatment and handling, including
ventilation and monitoring equipment. Each source ?bre is coupled with a single detector
?bre bundle using a connector which holds both in contact with the surface of the patient or
phantom. Each connector provides a common circular illumination and detection area with
a diameter of 6 mm, and is designed to facilitate the acquisition of re?ectance measurements
to calibrate the temporal characteristics of the imaging system (Hebden et al 2002b). The
interface which holds up to 32 connectors in contact with the head of an infant must also
optically isolate each detector from direct illumination by sources in neighbouring connectors
and from all external sources of light. Further design considerations are the requirement of the
reconstruction algorithm for precise 3D locations of every source and detector, and for these
locations to remain ?xed throughout the imaging scan.
While alternative designs for a single interface which can accommodate a variety of head
shapes and sizes over a broad range of gestational ages are being considered, initial studies
have utilized interfaces custom-built for each individual infant in the form of a plastic, foam-
lined helmet which ?ts over the back and top of the head. The shape and dimensions of
the helmet are based on a series of measurements acquired from digital photographs of the
infant taken one or two days prior to the study. The outer shell of the helmet is constructed in
two halves from low-temperature thermoplastic, and is lined with a soft NIR-absorbing foam,
10 mm thick, which is pre-sterilized using a low-temperature gas-plasma sterilization process.
The lower half, which ?ts beneath the rear hemisphere of the head, supports the weight of the
head and is held a few centimetres above the cot by a plastic frame. The top section of the
helmet ?ts over the forehead, and is supported by three vertical rods. The rods allow the top
section to be translated vertically and ?xed at any known position, while also enabling the

---

Three-dimensional optical tomography of the premature infant brain
4159
Figure 1. A homogenous reference phantom in?ated within a helmet used for optical tomography
of the infant brain.
infant to be instantly released from the helmet if required. The source/detector connectors
are attached to the helmet via small sockets mounted on the thermoplastic shell. The sockets
enable the connectors to be translated radially through holes in the foam lining towards the
skin and ?xed in position after the infant is placed in the helmet.
2.3. The reference phantom
The TOAST reconstruction algorithm is designed to reconstruct images of the absolute
values of the internal absorption and scattering coef?cients from sets of time-resolved
measurements recorded between ?bres placed at well-de?ned positions on the surface. In
practice, the positions of the sources and detectors are measured accurately with a 3D
digitizer (MicroScribe 3D, Immersion Corporation, USA) immediately before or after a clinical
measurement. However, since the helmet structure is necessarily mechanically deformable,
the sources/detectors inevitably move when the infant is placed in the helmet, and the recorded
positions are somewhat unreliable. Measuring the positions during the scan is impractical
since the infant’s head prevents access to the rear section of the helmet. While reconstruction
of absolute images is very sensitive to error in the source/detector positions, computer
simulations indicate that the requirement for accurate positions becomes signi?cantly less
stringent when reconstructing images using differences in data resulting from a change in the
optical properties of the head. In principle, if a change is measured relative to a reference
state with precisely known properties, absolute images can still be generated. We have
provided a reference measurement by using a homogeneous phantom inserted into the helmet
immediately following an infant scan. The phantom consists of a balloon ?lled with a solution
of intralipid and near-infrared dye, with a transport scattering coef?cient µ of 1.0 mm?1 and an
s
absorption coef?cient µa of 0.01 mm?1. Figure 1 shows the reference phantom placed within a
helmet.

---

4160
J C Hebden et al
Figure 2. A ?bre holder helmet on the head of a newborn premature infant during an imaging
scan.
3. Clinical measurements
MONSTIR was transferred to the UCL neonatal intensive care unit for the ?rst time in
October 2001. Initial studies on two unventilated premature infants enabled the helmet design
to be evaluated and modi?ed, and established that it was suf?ciently comfortable to be worn
continuously for several hours without interfering with the normal handling of the infant.
Figure 2 shows the helmet attached to the head of a 10-day-old infant born after 30 weeks
gestation. For a third study, the reference phantom was employed for the ?rst time, enabling
us to attempt an image reconstruction using difference data. The study involved a 5-week-old
male infant, born after 30 weeks gestation, who had suffered from a perinatal haemorrhage.
An ultrasound scan (see ?gure 3) revealed a large left intraventricular haemorrhage (grade 3)
and a much smaller right intraventricular haemorrhage (grade 2). The left ventricle is ?lled
and distended with blood, while the right ventricle contains blood but is not distended. There
is no evidence of haemorrhage in the brain parenchyma. The helmet contained 31 sources
and detectors distributed over the back and top of the head. Data acquisition proceeded
automatically as each source was illuminated for 15 s, and TPSFs were recorded by each
detector simultaneously. The total source intensity was within the IEC recommended limits
for both eye and skin safety. Each scan required about 9 min. Calibration measurements
were also acquired, which involved detecting light back-re?ected at the skin surface while
the source illumination was heavily attenuated. These measurements, and prior knowledge
of the temporal characteristics of each source ?bre, enable a full temporal calibration of the
data to be performed (Hebden et al 2002b). Two full sets of data were acquired on the infant,
followed by an identical measurement on the reference phantom in?ated to exactly ?ll the
helmet after the infant had been carefully removed. Locations of the sources/detectors on the
helmet were measured the following day using a 3D digitizer.
In order to generate 3D images of the internal optical properties, TOAST requires an
FEM model of the infant head with a realistic geometry. To provide a suitable mesh, we

---

Three-dimensional optical tomography of the premature infant brain
4161
Figure 3. Ultrasound image of the infant brain, with boundaries of the haemorrhage in both
ventricles indicated. The image represents a coronal slice with the top corresponding to the
location of the anterior fontanelle.
?rst acquired a 3D CT-scan of a realistic doll’s head, from which we generated a surface
mesh. Appropriate software was then used to apply a nonlinear warp to the surface mesh in
order to ?t it to the measured locations of the sources and detectors on the helmet. Finally,
the resulting surface was used to construct a volume mesh, containing 17 559 second-order
tetrahedral elements having a total of 26 814 nodes.
TOAST accommodates data in the form of speci?c datatypes, which represent certain
characteristics of each measured TPSF, such as temporal moments.
Simulations and
previous experimental investigations have indicated that a particularly effective combination
of datatypes is mean ?ight time and total intensity. However, the uncertain and variable
coupling at the surface (caused to a large degree by hair) currently prevents the routine use of
intensity, although methods are being explored to incorporate the determination of coupling
coef?cients into the reconstruction process (Boas et al 2001b). Temporal variance has also
been used effectively as a datatype although in this case variance values were too noisy due to
relatively poor photon statistics. Consequently our ?rst attempt to construct a 3D image of a
neonatal head was performed using measurements of mean ?ight time only, or, more precisely,
using differences between the mean ?ight times measured for the head and the corresponding
values for the intralipid ?lled balloon. TOAST performed 25 iterations using the data recorded
at each wavelength, starting from a homogenous estimate with properties corresponding to
those of the reference phantom. Each iteration required 25 min on a 2.2 GHz PC with 2 Gb
of RAM.

---

4162
J C Hebden et al
Figure 4. Transverse, coronal, and sagittal slices across the 3D absorption and scatter images
generated at a wavelength of 780 nm.
Figure 5. Transverse, coronal, and sagittal slices across the 3D absorption and scatter images
generated at a wavelength of 815 nm.
4. Results
Figures 4 and 5 show transverse, coronal and sagittal views across the 3D images of the
infant head, representing the distribution of absorption and scatter at the 780 nm and 815 nm

---

Three-dimensional optical tomography of the premature infant brain
4163
Figure 6. Transverse, coronal, and sagittal slices across the 3D images of estimated blood volume
and fractional oxygen saturation.
wavelengths, respectively. The sagittal view corresponds to the midplane, and the coronal
and transverse views are centred on the expected location of the cerebral ventricles. The
use of a single datatype for the reconstruction inhibits good separation between absorption
and scatter, and some degree of crosstalk is inevitable. Previous experiments on phantoms
have indicated that absorption images are generally more reliable than scatter images under
these circumstances, and this appears to be the case here. The coronal images clearly exhibit
greater absorption on the left side of the brain, consistent with the larger intraventicular
haemorrhage. The transverse and sagittal images at both wavelengths are dominated by a
feature at the rear of the head. While errors in the data or in the reconstruction process cannot
be ruled out as the principal origin of the feature, the sagittal sinus is a possible cause, since
it represents a dominant source of venous blood close to the surface. The distributions of
scatter at both wavelengths are clearly artefactual, partly as a consequence of crosstalk, and
partly due to the lower inherent tolerance of scatter images to the limitations of the forward
model. Both absorption and scatter images will be in?uenced by the inability of the forward
model to correctly represent the propagation of light through the non-scattering regions of
cerebral-spinal ?uid (CSF) which normally surround the brain and ?ll the ventricles.
By making an (albeit simplistic) assumption that the oxygenated and de-oxygenated
forms of haemoglobin are the only wavelength-dependent chromophores, the absorption
images at both wavelengths can be combined to generate images which display the total
volume of haemoglobin and the fractional oxygenation.
The images shown in ?gure 6
were generated using the known extinction coef?cients of both forms of haemoglobin, a
haemoglobin concentration in whole blood of 1.58 mM derived from venous samples taken
from the infant, and a ?xed value of 0.003 mm?1 for the background absorption (see Hillman
(2002) for a description of the method). The background value is based on an estimate by

---

4164
J C Hebden et al
van der Zee (1993) that other chromophores contribute about 30% of the total absorption in
the newborn infant head.
The volume image has a form very similar to both absorption images, and therefore
exhibits the expected asymmetry due to the haemorrhage and the strong feature at the rear
of the head. The saturation image is very nearly uniform, apart from a small region of very
low saturation on the front of the head, which is probably artefactual. Despite the simplistic
assumptions involved, the reconstructed values of both parameters agree reasonably well with
those expected for an infant head. NIRS measurements of absolute cerebral haemoglobin
concentrations in neonates by Wolf et al (1997) found an average tissue saturation of about
66 ± 4%, while we obtain a somewhat lower value of 54 ± 12% averaged over the top
half of the infant head. Meanwhile Wyatt et al (1990) used NIRS of the infant brain to
derive a mean value of cerebral blood volume of 2.2 ± 0.4 ml per 100 g of tissue, which is
consistent with our blood volume average of 2.3 ± 0.7%. This corresponds to an estimated
haemoglobin concentration of 36 ± 11 µM of tissue. Note that some differences between
total and partial volume measurements are inevitable due to the inclusion of regions which do
not include blood, such as those occupied by CSF. A partial volume effect due to the ?nite
spatial resolution prevents the haemorrhage from being exhibited as a region of 100% blood
volume, although these regions are obviously at the full concentration of blood. It is also
important to be aware that previous work with tissue-equivalent phantoms has demonstrated
that the limitations of the forward model as well as various sources of noise can signi?cantly
effect the reliability of quantitative values derived from reconstructed images, and therefore
we suggest the values presented here should be treated with some caution.
5. Discussion
Although alternative ‘one-size-?ts-all’ solutions are still being investigated, the design of
the custom-made helmets has proven to be both clinically acceptable and highly effective.
However, a current lack of knowledge of the surface geometry of the infant head and of the
precise positions of the source/detector locations during the scan still represents a problem for
image reconstruction. We are investigating the use of photogrammetry techniques to measure
the surface (as utilized by Bluestone et al (2001) for optical imaging of the adult forehead), and
alternative technologies are also being explored to provide real-time measurements of ?bre
locations. Meanwhile, so-called difference imaging using the adaptable reference phantom
at least partially makes up for the lack of precise geometrical information, although we are
currently performing a more detailed investigation into the effectiveness of this approach, and
its potential for producing image artefacts.
Despite some unwelcome (but currently unavoidable) compromises involved in the
reconstruction of the infant brain images, the results are promising. The absorption and
blood volume images exhibit the expected asymmetry due to the larger haemorrhage in the
left ventricle, and with the exception of the scatter images which are clearly dominated by
artefacts, the absolute values of tissue coef?cients and derived physiological parameters are
generally sensible. The extent to which CSF contributes towards the dominant C-shaped
region of low scatter exhibited across the central sagittal plane of the 3D scatter images
requires further investigation, and appropriate experiments on phantoms are currently in
progress. Improvement in both image quality and quantitative accuracy requires optimal
extraction of information from the time-resolved data (i.e. to utilize more than just mean
?ight time), and the implementation of forward models able to accommodate all the internal
structures and optical properties likely to be encountered within the infant brain. Since
TOAST can only reconstruct what the forward model is capable of representing, signi?cantly

---