Child’s Nerv Syst (1998) 14: 322–327
© Springer-Verlag 1998

ORIGINAL PAPER

Spiros Sgouros
Kalyan Natarajan
A. Richard Walsh
Edward B. Rolfe
Anthony D. Hockley

Computer simulation of a neurosurgical
operation: craniotomy for hypothalamic
hamartoma

Received: 8 January 1998

S. Sgouros (½) · A. R. Walsh
A. D. Hockley
Department of Neurosurgery,
Birmingham Children’s Hospital,
Steelhouse Lane,
Birmingham B4 6NH, UK
Tel.: +44-121-333 8146
Fax: +44-121-333 8151
K. Natarajan
Department of Medical Physics,
Birmingham Neuroscience Centre,
Queen Elizabeth Hospital,
Birmingham, UK
E. B. Rolfe
Department of Neuroradiology,
Birmingham Neuroscience Centre,
Queen Elizabeth Hospital,
Birmingham, UK

Abstract Although magnetic resonance imaging has revolutionised the
management of intracranial lesions
with improved visualisation of anatomical structures, it only produces
two-dimensional images, from which
the clinician has to extrapolate a
three-dimensional interpretation.
Several approaches can be used to
create 3D images; the discipline of
image segmentation has encompassed a number of these techniques.
Such techniques allow the clinician
to delineate areas of interest. The resulting computer-generated outlines
can be reconstructed in a three-dimensional arrangement. Although a
plethora of “generic” segmentation
techniques exist, we have developed
a refined form, dependent on general
and particular properties of the anatomical structures under investigation. High-contrast structures such as
the ventricles and external surface of
the head are found by using a localised adaptive thresholding technique.
Less definable structures, with poor
or nonexistent signal change across
neighbouring structures, such as
brain stem or pituitary, are found by
applying an “energy minimisation”based technique. To demonstrate the
techniques we used the example of

an 8-year-old boy with uncontrolled
gelastic seizures due to a hypothalamic hamartoma, who is being considered for surgery. We were able to
demonstrate the anatomical relationships between the hypothalamic hamartoma and adjacent structures
such as optic chiasm, brain stem and
ventricular system. We were subsequently able to create a video, reproducing the stages of craniotomy for
excision of this tumour. By creating
true 3D objects, we were able at any
stage of the simulation to visualise
structures situated contralaterally to
the approaching surgical dissector.
These 3D representations of the
structures can be either invisible or
opaque, in order to afford 3D localisation as the “virtual” surgical dissection proceeds. The clinical application of such techniques will enable
surgeons to improve their understanding of anatomical relations of
intracranial lesions and has obvious
implications in image-guided surgery.
Key words Magnetic resonance
imaging · Medical computing ·
Image-guided surgery · Image
segmentation · Computer graphics

323

Fig. 1 T1-weighted contrast-enhanced MRI scan (a sagittal, b coronal view) of an 8-year-old boy with intractable gelastic seizures. A
hypothalamic hamartoma is seen arising from the inferior aspect of
the right hypothalamus, anteriorly to the brain stem. The tumour is
well defined in all its aspects apart from the area of origin, where the
division from normal hypothalamus becomes indistinct. The lesion
is not enhancing with contrast. The optic chiasm has been shifted by
the tumour superiorly and anteriorly

Introduction

The advent of magnetic resonance imaging (MRI) has undoubtedly revolutionised the management of intracranial
pathology. While MRI scanning offers improved detail on
soft tissues, in its current form it produces only two-dimensional views of the pathology under scrutiny. Neurosurgeons have to mentally re-create a 3D reconstruction by
extrapolating from 2D images while planning an operation. Any means of providing 3-dimensional graphical representations would help surgeons enhance their understanding of anatomical relationships of deep-seated structures and advance intracranial surgery even further.
Owing to differences in the digital signal and the ranges
that MR imaging parameters (TR, TE) can take, the simple thresholding-based techniques used to create 3D bone
and ventricular CSF reconstructions in CT scanning cannot be directly applied to MRI [4]. The variation of the
grey range of the MRI signal is such that simple thresholding would eradicate a significant number of intracranial
structures from the reconstruction, resulting in inaccuracies.
During the last decade, different segmentation techniques have been developed to overcome these difficulties
[4, 5, 11, 19, 20]. Image segmentation is the technique of
fragmenting a composite picture into several user-defined
outlines/regions, thus accurately defining real-world objects. Several algorithms can be used to achieve this, ranging from manual to fully automatic methods [1, 2, 4, 17–20,

22]. The level of operator-independence achieved is related
to the accuracy and the level of structural complexity that
can be handled. The resulting outlines can be reconstructed
in 3D. Various image modalities can be used with these
techniques, including CT, MRI, MRA, SPECT and PET
scan.
Although the technique has been available for a while,
clinical application has been rather slow, owing partly to
the unfamiliarity of clinicians with computer technology
and to the time involved in achieving accurate and meaningful reconstructions. We demonstrate in this communication the clinical use of our segmentation approaches simulating the stages of craniotomy for intracerebral tumour.
The rare and surgically difficult-to-treat hypothalamic hamartoma served as a good example of the clinical value of
segmentation in creating 3D MRI images.

Materials and methods
An 8-year-old boy had been suffering from gelastic seizures for several years, lately with increasing frequency and refractory to medical treatment. He was referred for assessment of suitability for surgical treatment.
On MRI scanning he was found to have a mostly well-defined,
non-contrast-enhancing lesion arising from the right hypothalamus,
which was thought to be a hypothalamic hamartoma (Fig. 1). At the
area of origin from the hypothalamus, the definition between normal
tissue and hamartoma was indistinct. The MRI examination used for
the reconstruction was performed on an IGE 1.5 T scanner (IGE, Milwaukee, USA) and consisted of 124 contiguous images, T1-weighted (TE:20/TR:200), of 1.5 mm thickness, obtained following intravenous injection of paramagnetic contrast.
The availability of 3D MRI images would be particularly beneficial in the surgical planning in the case of such a deep-seated lesion. The surgical approach has several associated technical difficulties, mainly related to the anatomical relationship of the lesion
to important structures such as the brain stem, the optic chiasm,
and the pituitary stalk. 3D reconstruction would improve spatial
perception of the lesion and assist surgical planning of a direct approach.

324

To perform the segmentation task we used an in-house X-Windows-based C program running on a UNIX operating system workstation. The workstation, a Sun SparcStation 20 (Sun Computers,
California), is connected via a local area network to the MRI and CT
scanner systems. The software environment has been developed with
the ability to handle numerous image modalities and formats. Additionally, images obtained in other scanners can be imported via external media such as DAT, optical disc and floppy disk.
The software incorporates segmentation algorithms and 3D visualisation routines. Automatic, semiautomatic or manual algorithms
can be used to outline structures, depending on the complexity of
their shape and the difference in grey value from adjacent structures
[5, 6, 15, 16].
In the case presented here, automatic segmentation was used for
the skin and the cortical surface. The skin was easily outlined automatically, as the air–skin interface is clearly identifiable as a highcontrast area. For segmentation of the cortical surface CSF boundaries were determined from a seed derived from the centre of gravity
of the current scan. This seed point was subsequently used for the
other slices. This information was recorded in a database of cortical
outlines for later use. When this technique failed a more “user-involved” semiautomatic algorithm was employed. Constraints on grey
level, first- and second-order derivatives were used as primary factors [6]. The constraints were augmented by energy minimisation
technique, adjusting the contrast between neighbouring structures,
to isolate the region of interest. Where the boundary was weak or
nonexistent, a constraining boundary of arbitrary shape was applied
around the area of interest, to improve the possibility of semiautomatic delineation. The arbitrary shape was passed on to adjacent slices, to follow anatomical continuity of the structure of interest. This
technique worked well for the ventricular system, the brain stem and
the chiasm. The segmentation program allows a range of slices to be
user-defined, with their anterior and posterior extent determined on
the mid-sagittal plane. This enhances the possibility of correct results with semiautomatic methods. Whenever the boundaries were
of the correct shape but incorrect size, too big or too small, the techniques of mathematical morphology erosion and dilatation were employed to improve the result. Finally, if errors still persisted, the contour was edited manually. A typical example was the hypothalamic
tumour. The ill-defined plane of origin from the inferior aspect of
hypothalamus had to be manually outlined. For selected objects, volumes were measured automatically from cross-sectional area and
slice thickness.
The segmentation is performed on 2D images. The anatomical
outlines produced are reconstructed in 3D. For the high-resolution
dataset the volume is constructed by summing the cross-sectional
area. For those cases where the 3D resolution is poor, volumes are
generated by interpolation between cross sections. These object
volumes are then superimposed on the actual 3D MRI data. Each
structure outlined becomes a separate 3D object, coloured individually to be distinguishable. The surface geometry of the MRI data,
for example the cortex, can then be viewed together with the
coloured outline objects. Two modes of visualisation exist: depth
shadows and perspective shading. These objects can be made transparent or removed completely, allowing demonstration of deeper
situated structures. They can also be rotated and viewed from any
desired angle. Simulated “surgical” cuts can be made in any desired trajectory, allowing customised visualisation. Subsequently a
video can be produced, in predetermined sequences, simulating the
entire virtual “operation”. The whole procedure of 3-dimensional
reconstruction is performed by the neurosurgeon and can take from
15 to 60 min, depending on the number of structures that are outlined, their complexity and their MR signal appearance.
The environment described above has been in daily clinical use
for several years to measure hippocampal volumes in epileptic patients scheduled for hippocampectomy. In the last 2 years its use has
been extended to surgical simulation of tumour resections and craniofacial research.

Results

Using mostly automatic and semiautomatic outlining, after 40 min of computing work, a complete 3D reconstruction of the child’s head had been achieved, including skin
with marked skin flap, cerebral hemispheres, ventricular
system, brain stem, tumour, and optic chiasm. Figure 2
demonstrates the process of segmentation. The areas of
interest have been outlined. The resulting outlines were
subsequently reconstructed in 3D together with the real
MR digital data.
For this demonstration suitable viewpoint 3D co-ordinates were chosen to simulate the surgeon’s view of the
patient’s head on the operating table. Figures 3–8 are stills
from a video sequence simulating the steps of a craniotomy for resection of this tumour. In Fig. 3 the skin surface
is seen, with the proposed skin flap marked in a different
colour (purple). In Fig. 4 the skin flap is made transparent
to demonstrate the relation of the flap to the underlying
sylvian fissure. In Fig. 5 the superficial part of the brain
has been made transparent to demonstrate the relationship
of the tumour to the brain stem (green). This relationship
is better demonstrated in Fig. 6, where the right frontal lobe
has been removed, exposing the ventricular system (purple), the tumour (red), and the optic chiasm (yellow). In
Fig. 7 an additional surgical approach via the subfrontal
route is simulated, by “removing” the anterior part of the
frontal lobe through a frontal “craniotomy”. In Fig. 8 all
the superficial structures have been removed to expose the
tumour in relation to the ventricular system and the chiasm, which is seen to be severely deformed by the underlying tumour. The ability to rotate the 3D object allows better visualisation of the relationship of the tumour to the
surrounding important structures. The tumour volume has
been calculated as 3913 mm3.

Discussion

Three-dimensional visualisation of intracranial structures
offers realistic perception of the area of interest and, consequently, improved understanding of anatomical relationships. The high level of soft tissue detail that MRI offers
combined with the stereoscopic visualisation of 3D reconstruction are opening a new perspective in surgical planning of deep intracranial lesions. The ability to interact
with 3D structures and visualise their anatomical relationships from multiple viewpoints is helping surgeons rapidly
increase their own personal experience.
Segmentation techniques have been present for some
time now, and most of them have been used in various institutions predominantly for research [3, 12, 13]. The presence of in-house computer expertise allows the flexibility
of continuing development according to the clinical need.
Recently, simple forms of segmentation programs have ap-

325

Fig. 2 Segmentation of a composite image. Using automatic,
semi-automatic or manual algorithms, the composite MRI image is “split” into simple multiple outlines of the structures of
interest. Ventricular system
(blue line), optic chiasm
(yellow line), tumour (red line)
and pituitary stalk (brown line)
have been outlined
Fig. 3 3D reconstruction of
the patient’s head. The skin flap
(purple) is marked in the region
of the right pterion
Fig. 4 The skin flap is made
transparent, allowing visualisation of the posterior aspect of
the underlying sylvian fissure

2

3

4

5

6

7

8a

b

Fig. 5 The superficial part of
the brain is made transparent.
The surgical approach to the tumour (red) can be verified
against deep-seated structures
(lateral ventricle purple, brain
stem green)
Fig. 6 The right frontal lobe
has been removed, to expose
the ventricular system (purple),
the optic chiasm (yellow), and
the tumour (red). The displacement of the optic chiasm by the
tumour is observed
Fig. 7 In addition to the pterional approach, a subfrontal approach has been simulated by
removing the anterior part of
the frontal bone and the frontal
lobe. The relationship of the tumour to surrounding deep-seated structures is demonstrated
(ventricular system purple,
brain stem green, tumour red)
Fig. 8 a Lateral and b oblique
views when all the superficial
structures have been removed
to visualise the tumour (red) in
relation to the ventricular
system (purple) and the optic
chiasm (yellow). The optic chiasm displacement superiorly
and anteriorly is clearly appreciated. Rotation of the three-dimensional object allows detailed assessment

326

peared as part of image manipulation software in imageguided surgery systems.
Widespread clinical application of image segmentation
has been limited so far. This is largely due to the unfamiliarity of most neurosurgeons with computing, and the time
and the level of interaction involved to create 3D objects
of satisfactory quality for clinical use. Both factors operate synergistically, creating a gap between surgeons and
computers. The busy clinical schedule of most neurosurgeons does not allow the luxury of a few spare hours to experiment with 3D MRI.
With improving technology both these factors are likely
to change. The computer is gradually infiltrating into neurosurgical practice, initially as part of frameless stereotaxy
systems. Parallel improvements in MRI technology and advances in computers’ memory capacity and speed of execution are rapidly reducing the time required for 3D and
4D (including time axis) reconstructions.
Several workers have attempted to make use of these
valuable segmentation techniques in clinical practice.
Fields such as skull base or craniofacial surgery lend themselves particularly well to this purpose. Although it has not
taken off routinely as yet, the feasibility of surgical planning using 3D reconstruction has been demonstrated, several imaging modalities often being combined [1, 3, 8–10,
13, 14, 23, 25].
Several fields of neurosurgery are likely to benefit from
this technology. Brain tumour surgery, assisted by image
guidance and 3D MRI, would become safer and more radical. Follow-up of cerebral tumours has obvious benefits
from volumetric serial studies [17]. Stereotactic radiosur-

gery and epilepsy surgery would benefit from any 3D volumetric imaging technology [2]. In addition to surgical
planning, surgical teaching would be improved by the use
of 3D MRI. The advent of 3D atlases of the head [21, 24]
is promising to revolutionise neurosurgical teaching. The
difference between idealised atlases and a true 3D MRI object is that the latter conforms accurately to reality, at least
in its preoperative imaging. In a similar way, the presented
simulation of a neurosurgical operation would allow the
surgeon to visualise the true anatomical relationships of
the various structures to one another and “perform” the operation on the computer, thus anticipating potential surgical difficulties.
As the rate-limiting step remains the operator dependence and the associated time factor, the obvious next steps
for further development would be in the area of improved
automation. The incorporation of physics, anatomical and
continuity constraints, which are currently under development in our laboratory [15, 16], would make the technique
more appealing and will help establish its routine clinical
use. Several workers have contributed significantly on these
lines in recent years [1, 2, 4, 13, 17, 18, 22, 23]. Further development on dynamic MRI acquisition [7] is likely to improve the chances of existing automatic segmentation by
improving the definition between various brain structures,
which share similar grey values in ordinary MR sequences.
With wider use of segmentation in image-guided
systems and the development of tele-surgery, a new perspective in visualisation of neurosurgical pathology is
likely to evolve, which with time will improve surgical
techniques.

References
1. Ardekani BA, Braun M, Hutton BF,
Kanno I, Iida H (1995) A fully automatic multimodality image registration
algorithm. J Comput Assist Tomogr
19: 615–623
2. Ashton EA, Berg MJ, Parker KJ, Weisberg J, Chen CW, Ketonen L (1995)
Segmentation and feature extraction
techniques, with applications to MRI
head studies. Magn Reson Med
33: 670–677
3. Chalif DJ, Dufresne CR, Ransohoff J,
McCarthy JA (1988) Three-dimensional computed tomographic reconstructions of intracranial meningiomas.
Neurosurgery 23: 570–575
4. Clarke LP, Velthuizen RP, Camacho
MA, Heine JJ, Vaidyanathan M, Hall
LO, Thatcher RW, Silbiger ML (1995)
MRI segmentation: methods and applications. Magn Reson Imaging
13: 343–368

5. Davis DN, Natarajan K, Claridge E
(1994) Advanced segmentation. (AIM
project A2032) Commission for the
European Committees, DG 13
6. Davis DN, Natarajan K, Claridge E
(1995) Multiple energy function active
contours applied to CT and MR images. IEE Proceedings: Vision, Image
and Signal Processing 1995. IEE Conference, Publication 410: 114–116
7. Duyn JH, Mattay VS, Sexton RH,
Sobering GS, Barrios FA, Liu G, Frank
JA, Weinberger DR, Moonen CT
(1994) 3-Dimensional functional imaging of human brain using echo-shifted
FLASH MRI. Magn Reson Med
32: 150–155, 545 (erratum)
8. Ehricke H, Daiber G, Sonntag R,
Strasser W, Lochner M, Siemens AG,
Rudi LS, Lorenz WJ (1992) Interactive
3D graphics workstations in stereotaxy
– clinical requirements, algorithms and
solutions. Visualisation Biomed Comput SPIE 1808: 548–552

9. Gandhe AJ, Hill DL, Studholme C,
Hawkes DJ, Ruff CF, Cox TC, Gleeson
MJ, Strong AJ (1994) Combined and
three-dimensional rendered multimodal data for planning cranial base surgery: a prospective evaluation. Neurosurgery 35: 463–470
10. Hayashi N, Endo S, Kurimoto M,
Nishijo H, Ono T, Takaku A (1995)
Functional image-guided neurosurgical
simulation system using computerized
three-dimensional graphics and dipole
tracing. Neurosurgery 37: 694–703
11. Holden M, Steen E, Lundervold A
(1995) Segmentation and visualization
of brain lesions in multispectral magnetic resonance images. Comput Med
Imaging Graph 19: 171–183

327

12. Kikinis R, Cline HE, Altobelli D,
Halle MW, Lorensen WE, Jolesz FA
(1992) Interactive visualisation
and manipulation of 3D reconstructions for planning of surgical procedures. Visualisation Biomed Comput
SPIE 1808: 553–559
13. Kikinis R, Gleason PL, Moriarty TM,
Moore MR, Alexander E III, Stieg PE,
Matsumae M, Lorensen WE, Cline
HE, Black PMcL, Jolesz FA (1996)
Computer-assisted interactive three-dimensional planning for neurosurgical
procedures. Neurosurgery 38: 640–651
14. Matsumae M, Kikinis R, Mórocz I,
Lorenzo AV, Albert MS, Black PMcL,
Jolesz FA (1996) Intracranial compartment volumes in patients with enlarged ventricles assessed by magnetic
resonance-based image processing.
J Neurosurg 84: 972–981
15. Natarajan K (1993) A knowledge
based system for radiological image
recognition. PhD thesis, University of
Birmingham

16. Natarajan K, Claridge E, Davis DN
(1996) The symbolic atlas of the brain:
handling non-visual information associated with neuroanatomy. Med Inf
(Lond) 21: 1–21
17. Phillips WE II, Velthuizen RP, Phuphanich S, Hall LO, Clarke LP, Silbiger ML (1995) Application of fuzzy
c-means segmentation technique for
tissue differentiation in MR images of
a hemorrhagic glioblastoma multiforme. Magn Reson Imaging
13: 277–290
18. Porrill J, Ivins J (1994) A semiautomatic tool for 3D medical image analysis using active contour models. Med
Inf (Lond) 19: 81–90
19. Robb RA, Barillot C (1989) Interactive display and analysis of 3D medical images. IEEE Trans Biomed Eng
8: 217–226
20. Robb RA, Hanson DP (1991) A software system for interactive and quantitative visualization of multidimensional biomedical images. Australas Phys
Eng Sci Med 14: 9–29
21. Schubert R, Hohne KH, Pommert A,
Riemer M, Schiemann T, Tiede U,
Lierse W (1994) A new method for
practicing exploration, dissection, and
simulation with a complete computerized three-dimensional model of the
brain and skull. Acta Anat (Basel)
150: 69–74

22. Sivewright GJ, Elliott PJ (1994) Interactive region and volume growing for
segmenting volumes in MR and CT
images. Med Inf (Lond) 19: 71–80
23. Sundaramoorthy G, Hoford JD,
Hoffman EA, Higgins WE (1995)
IMPROMPTU: a system for automatic
3D medical image-analysis. Comput
Med Imaging Graph 19: 131–143
24. Tiede U, Bomans M, Höhne KH,
Pommert A, Riemer M, Schiemann T,
Schubert R, Lierse W (1993) A computerized three-dimensional atlas of
the human skull and brain. AJNR Am
J Neuroradiol 14: 551–559
25. Valentino DJ, Mazziota JC, Huang HK
(1991) Volume rendering of multimodal images: application to MRI and PET
imaging of the human brain. IEEE
Trans Biomed Eng 10: 554–562