Digitisation and 3D Reconstruction of 30 Year Old
Microscopic Sections of Human Embryo, Foetus
and Orbit
Joris E. van Zwieten1, Charl P. Botha1, Ben Willekens2,
Sander Schutte3, Frits H. Post1, and Huib J. Simonsz4
1 Data Visualisation Group, Delft University of Technology
2 The Netherlands Ophthalmic Research Institute, Amsterdam
3 Biomechanical Engineering, Delft University of Technology
4 Ophthalmology Department, Erasmus Medical Centre, Rotterdam
Abstract. A collection of 2200 microscopic sections was recently recovered at
the Netherlands Ophthalmic Research Institute and the Department of Anatomy
and Embryology of the Academic Medical Centre in Amsterdam. The sections
were created thirty years ago and constitute the largest and most detailed study of
human orbital anatomy to date. In order to preserve the collection, it was digitised.
This paper documents a practical approach to the automatic reconstruction of a 3-
D representation of the original objects from the digitised sections. To illustrate
the results of our approach, we show a multi-planar reconstruction and a 3-D
direct volume rendering of a reconstructed foetal head.
1 Introduction
Recently, a collection of 2200 microscopic sections from 2 human embryos, 7 human
foetuses, and 3 adult orbits was recovered at the Netherlands’ Ophthalmic Research
Institute and the Department of Anatomy and Embryology of the Academic Medical
Centre in Amsterdam.
The orbit is the bony cavity that contains the eye and its appendages, i.e. the eye
socket. A microscopic section is a very thin slice of tissue, laid flat on a glass slide,
stained, mounted in a medium of proper refractive index and covered with a very thin
piece of glass called a coverslip [1]. Figure 1 shows an example of two microscopic
sections of a 79 mm (crown to rump length) human fetus from the recovered collection.
The collection, henceforth referred to as the orbita collection, forms the heritage of
the late J.A. Los, L. Koornneef, and A.B. de Haan. The sections were created between
1972 and 1986. Koornneef studied the development and the adult configuration of pre-
viously unknown connective tissue septa, or thin partitions, found within the orbital
fat. De Haan studied the development of the bones that compose the orbit. Especially
the work by Los and Koornneef has become well known among anatomists worldwide.
The collection of sections itself can be considered one of the largest and most detailed
studies of orbital anatomy to date.
When the sections were recovered it was noticed that the dyes used for staining had
started to fade. To preserve the collection, all the sections were digitised at an optical
A. Campilho and M. Kamel (Eds.): ICIAR 2006, LNCS 4142, pp. 636–647, 2006.
c© Springer-Verlag Berlin Heidelberg 2006

638 J.E. van Zwieten et al.
of each diascopic image onto its corresponding episcopic image guarantees that the
anatomical differences between adjacent sections are preserved. This is an important
advantage over methods that apply non-rigid registration to pairs of adjacent diascopic
sections. Additionally, registration error only affects a single section and does not prop-
agate through the stack.
Unfortunately, no episcopic images were recorded when the orbita collection was
created. In this case, the canonical approach is to register pairs of adjacent sections
[8,9,10]. The authors of [10] cluster the displacement vector field and estimate an affine
transform for each cluster, which are subsequently embedded into one non-rigid trans-
form. Their approach is to tailor the transform space to the observed deformation pat-
tern. This helps to avoid transforms that are known to be wrong from the outset. Instead
of performing registration on pairs of adjacent sections, some authors register a section
with respect to both adjacent sections [11,12]. This is an interesting method, because it
encourages transforms that yield smooth structures. If the original structures are known
to be smooth with respect to the section thickness, one can assume that rapid changes
in the direction of structures are a consequence of histological processing and should
therefore be suppressed.
3 Preparation of Microscopic Sections
To study a specimen the size of a human orbit, embryo, or foetus under a light micro-
scope it has to be sectioned, i.e. cut into thin sections. The field of microtechnique is
concerned with the preparation of biological material for microscopic observation. This
section provides a short introduction to the field and its terminology (derived from [13])
as well as a description of the specific preparation of the orbita sections.
The creation of microscopic sections from a biological specimen can be divided into
five steps: fixation, embedding, sectioning, staining and mounting. The aim of fixation
is to terminate life processes quickly and preserve the organisation of cells with as little
deformation as possible. The canonical fixative is formaldehyde, which was also used
to fixate the specimens that compose the orbita collection.
A specimen is embedded to support it during sectioning. Three well-known em-
bedding methods are: paraffin embedding, nitrocellulose embedding, and freezing, of
which paraffin embedding is the most common. The orbita collection was embedded in
nitrocellulose.
Historically, the sectioning of a specimen was performed by hand using a razor.
Much better results can be obtained by using a microtome. Microtomes are machines in
which a block of embedded tissue and a knife can be fixed. To cut a section, either the
knife is drawn across the embedding block, or the block is drawn over the knife.
The various constituents of tissue (e.g. cytoplasm, connective tissue, and lipids) con-
tained in a section can be discriminated by staining. Dyes used for staining are often
applied in pairs: a basic dye, which stains constituents with predominantly acid group-
ings (also termed stain), and an acid dye (also termed counterstain).
Several combinations of dyes were used to stain the orbita collection sections (see
figure 1). Koornneef used haematoxylin-azophloxin,while De Haan used haematoxylin-
eosin (H-E). Both used haematoxylin-Van Gieson’s picro-fuchsin. We will discuss these

Digitisation and 3D Reconstruction of 30 Year Old Microscopic Sections 639
dyes in short. Because azophloxin and eosin yield highly similar results, they can be
assumed identical for our purposes.
Alum haematoxylin is a basic dye. This dye stains nuclei deep blue or purplish, hya-
line cartilage and bone is stained light to medium blue. Eosin is a red acid dye, which
stains cytoplasm, muscle fibres, blood corpuscles and connective tissue. The combina-
tion of haematoxylin and eosin displays a general histological overview. Van Gieson’s
picro-fuchsin is an acid dye, which can be used as a counterstain for haematoxylin. It
stains collagenous connective tissue, muscle, and keratin. Applied to embryos it pro-
duces a good distinction between bone and cartilage.
A section is mounted to preserve it and to enhance the visibility of the structures it
contains. A section is laid down on a glass slide, a suitable mounting medium is applied,
and a coverglass is placed over the section.
4 Digitisation
Digitisation of the orbita sections was performed at the Netherlands Ophthalmic Re-
search Institute. Each glass slide was placed into a bracket that was custom made to fit
an ArtixScanTM 4500t film scanner from Microtek.
This scanner is designed specifically to scan photographic film and slides. It has an
optical resolution of 2500 dpi at a bit depth of 14 bits per colour. Scanners are often
used to scan objects that are already records of three components of colour information
and therefore might use narrow band filters. If a natural object such as a histological
section is scanned with such a scanner the reported RGB values will be inaccurate [14].
However, according to the specifications of the scanner, the employed filters are not of
the narrow band type.
5 Reconstruction
Reconstruction of 3-D datasets from the digitised microscopic sections consists of a
pre-processing phase and a registration phase. In the pre-processing phase, the digitised
images are prepared for registration. During the registration phase, pairs of adjacent
images are registered onto each other. In other words, for each pair a transformation
is sought that fits one image of the pair (the floating image) onto the other image (the
reference image). The result of reconstruction is a 3-D dataset.
5.1 Pre-processing
The pre-processing phase consists of resampling, segmentation, and an RGB to scalar
data conversion.
Resampling. Firstly, the resolution of the sections is decreased by resampling. Re-
construction of a resampled series takes less time, which facilitates fast prototyping.
Reconstruction of the resampled data can be interpreted as the first step of a progres-
sive refinement algorithm: the solution computed for the resampled data can be used
as a starting point for reconstruction of the high resolution data. All sections were re-
sampled to a standard width of 512 pixels while maintaining the original aspect ratio.

640 J.E. van Zwieten et al.
Resampling was performed with a lanczos kernel. This is a windowed sinc kernel that
is often used for high quality image resampling.
Segmentation. Segmentation is the process of separating an object of interest from
the background in an image. It is relevant to reconstruction, because a reconstruction
should only be based on the object of interest (and/or external fiducials). As a corollary:
if the background is largely homogeneous, segmentation may be unnecessary. However,
even if segmentation is unnecessary for accurate reconstruction, it may still speed up
computation because background pixels can be ignored.
The output of the segmentation procedure is a binary mask of the object. An initial
segmentation is performed by exploiting colour differences between the background
and the object of interest. Separation was found to be easier in HSV (Hue, Saturation,
Value) colour space than in RGB (Red, Green, Blue) colour space. A colour threshold is
derived from the histogram. We experimented with two automatic threshold selection
methods: the triangle method and the background symmetry method, see [15, pp.95-
96]. The first was used if the histogram contained a single background peak, the latter
if it contained both a background and a foreground peak. However, in our experience,
manually selection of a fixed threshold yields the best results.
The initial thresholding marks holes within the object as background. However, as
these holes are generated by the object they can be considered as features for recon-
struction. Therefore, the segmentation procedure was extended to close the holes in the
initial mask. First, binary dilation is applied to close the object’s outer contour. Next,
binary propagation is used to remove unconnected background artifacts. Finally, inver-
sion followed by binary propagation of the background and another inversion closes any
remaining holes in the mask. Figure 2 illustrates the various steps of this segmentation
procedure.
Fig. 2. From left to right, the five steps of the segmentation procedure: input section, thresholding
on hue, binary dilation to close contour, background artifact removal, filling the remaining holes
in the mask
Conversion to Scalar Data. An essential component of a registration algorithm is
the similarity metric which quantifies the goodness-of-fit or similarity of two adjacent
sections. Most of these metrics are designed for scalar data, i.e. one value per position.
Microscopic sections are treated with dyes to differentiate the anatomical structures
of interest and therefore colour provides information about the tissue comprising the
section. Colour information is inherently multi-valued instead of scalar. Existing litera-
ture on the registration of microscopic sections describe similarity metrics designed for

Digitisation and 3D Reconstruction of 30 Year Old Microscopic Sections 641
scalar data and apply these metrics to intensity images. Of course, intensity by itself
does provide information about anatomical structure, which may well be sufficient to
attain accurate registration. However, there are examples of applications where colour
information can be used effectively to enhance the result, e.g. edge detection.
Instead of extending a similarity metric to account for colour data, we decided to
reduce the dimensionality of the data to allow the use of similarity metrics designed for
scalar data while minimising the loss of information.
A standard conversion to greyscale can be viewed as a mapping:
f (r,g,b) = r3 +
g
3 +
b
3 = [r g b] · [
1
3
1
3
1
3 ]
T
A colour is mapped onto the dot product between that colour and the vector [ 13
1
3
1
3 ]
T
.
In geometrical terms, f projects the RGB colour space onto the line through 0 with
direction vector [ 13
1
3
1
3 ]
T (and scales the result with the length of this vector).
Of course, we can project the colour space onto a line in any given direction. In fact,
we can use Principal Component Analysis or PCA to find the direction that retains the
largest part of the total variance compared to all other directions. Table 1 summarises the
results of this approach. Notice how the blue component is almost completely ignored
for the eosin sections, because it is approximately constant. Also, although Van Gieson’s
dye displays several hues on the orbita sections, i.e. purple-red-orange-yellow, these
colours approximately lie along a line in the RGB colour space. For the Van Gieson
sections, both PCA and straightforward conversion to grey scale retain a large part of
the total variance. In general, however, mapping colours onto the first principal direction
will preserve more variance than conversion to grey scale. Therefore, this method of
scalar conversion is recommended.
Table 1. The retained variance as a percentage of the total variance and the axis corresponding to
either greyscale or the first principal component. We randomly selected 75 sections of each dye
combination from series S73345 as input data.
Stain Method Retained variance Direction Spherical variance
Eosin GREY 63.11% (σ = 3.51) 0.5774 0.5774 0.5774 N/A
Van Gieson GREY 93.57% (σ = 1.87) 0.5774 0.5774 0.5774 N/A
Eosin PCA 95.03% (σ = 1.83) 0.4473 0.8930 0.0507 0.0011
Van Gieson PCA 97.61% (σ = 0.59) 0.3939 0.6669 0.6325 0.0014
The spherical variance1 of the axes found by PCA is small (see table 1). In other
words, variance within a group of sections stained with a single dye combination (e.g.
variance in staining time and fading of the sections) does not have much effect on the
direction of the axis of largest variance. Therefore, the average direction could be used
to convert all the sections of a series that were stained with a single dye combination.
1 Spherical variance is a measure of the dispersion about a mean direction, which has a range of
[0,1] (see [16, pp.218,219]).

642 J.E. van Zwieten et al.
5.2 Registration
In the registration phase, pairs of digitised sections are transformed to fit onto each
other. The selected transformation family determines the degrees of freedom of the
transformation. A similarity metric is used to quantify the quality of the fit. An optimi-
sation method searches the space of transforms to find a transform that maximises the
similarity metric. In the following subsections, we discuss these three components with
the registration of the orbita collection in mind.
Transformation Family. As argued by [10, p.2] the selection of a transformation fam-
ily should be guided by a priori knowledge of the deformation caused by preparation. If
the transformation family is too constrained it cannot adequately express the deforma-
tion, if it is too general it may introduce solutions that are known to be incorrect.
The sections related to the study of Koornneef were originally created to investigate
the spatial relations between newly discovered septa and the bony boundaries of the or-
bit. However, the preparation of the sections is quite favourable for 3-D reconstruction.
Firstly, nitrocellulose was used as embedding medium. Deformation of material embed-
ded in nitrocellulose during sectioning is relatively small compared to commonly used
alternatives such as paraffin. Deformation during embedding is also less prominent be-
cause the procedure does not require heat. Secondly, decalcification was performed with
E.D.T.A., a procedure that while slower causes less damage than commonly used alter-
natives. Thirdly, relatively thick sections were cut to enhance the visibility of the septa.
Thicker sections generally deform less easily. Visual inspection of the sections revealed
a nearly circular eyeball, which supports the idea that the sections suffered little overall
deformation [17,13]. The sections associated with the study of De Haan were created
to investigate the development of the orbit. The material was also embedded in nitro-
cellulose. Although nitric acid was used instead of E.D.T.A. for decalcification, visual
inspection revealed only minor deformations [18].
Because section deformation should be minor given the preparation procedure we
decided to use rigid transforms for registration. More general transforms could remove
differences between sections that are a consequence of anatomy [9, p.30]. Furthermore,
after performing reconstruction based on rigid transforms it is easier to judge the extent
of non-rigid deformation in the sections.
Similarity Metric. A similarity metric should be chosen that reflects the physical rela-
tion between the image intensities2 [19]. However, for the orbita collection this relation
is unknown.
To investigate the relation between the image intensities we manually registered a
small number of selected section pairs from series S73345 and plotted a joint histogram
of the overlaying intensities (see figure 3). The sections in series S73345 are alternat-
ingly stained with haematoxylin-eosin and haematoxylin-Van Gieson’s dye. The his-
togram in figure 3 (left) was created from a pair of sections that are not directly adjacent
but one section apart. The figure therefore shows the relation between a pair of sections
2 In this paper, the word intensity refers to the value obtained by converting colour data to scalar
data using PCA.

Digitisation and 3D Reconstruction of 30 Year Old Microscopic Sections 643
stained with a single dye combination. We also investigated the relation between sec-
tion pairs stained with alternating dye combinations, see figure 3 (middle and right).
However, as the results were similar, we will focus on the first case.
Fig. 3. Joint histogram of intensities after manual registration: (left) a pair of same-dye sections
(H-E), (middle) a pair of sections stained with alternating dye combinations (H-E, H-Van Gieson),
(right) a pair of sections stained with alternating dye combinations (H-Van Gieson, H-E). Val-
ues in the joint histogram that lie further than ten standard deviations from the mean have been
clamped to provide a clearer picture.
From figure 3 (left) the intensities seem to approximate a linear relation. The corre-
lation coefficient squared (r2) equals 0.695 for the haematoxylin-eosin pair and 0.729
for the haematoxylin-Van Gieson’s dye pair. In other words, 68.9% and 72.6% of the
total variance can be explained by linear regression respectively. We repeated the same
procedure on a pair of sections from series S74161, which is stained exclusively with
haematoxylin-eosin, and found r2 = 0.595.
The strength of these linear relations is only moderate, especially for the section pair
from series S74161. Figure 4 shows the optimisation surfaces generated both by a simi-
larity metric based on normalised correlation and one based on mutual information for
this section pair. The normalised correlation metric implicitly assumes a linear intensity
relation, while the mutual information metric assumes a much more general, statistical
relation.
Both similarity metrics have a distinct minimum that approximately coincides with
the manual alignment. It appears that the linear intensity relation is strong enough to be
detected reliably by the normalised correlation metric.
The normalised correlation metric has a larger capture range than the mutual infor-
mation metric. Also, it yields a ’smoother’ optimisation surface, because it is computed
on the entire overlap domain while the mutual information metric is computed on a
random sample of points from the overlap domain [20]. Because a smooth optimisation
surface and a large capture range are both desirable properties for a similarity metric,
we decided to use the normalised correlation metric for section pairs stained with a
single dye combination.
Figure 1 shows a pair of adjacent sections stained with different dye combinations.
Two things are apparent from this figure. First of all, the two dye combinations mostly
differentiate the same tissues. Intuitively one would expect this to be advantageous for
registration. Secondly, although the two dye combinations produce different hues, the
pattern of colour saturation seems roughly the same.

644 J.E. van Zwieten et al.
-0.94
-0.92
-0.9
-0.88
-0.86
-0.84
-0.82
-0.8
-0.78
-3 -2 -1 0 1 2 3
si
m
ila
rit
y
angle
S74161_ncc_large_angle.map
(start 3.1416, end -3.1416, steps 30)
S74161_ncc_large.map
(start -50.0000, end 50.0000, steps 30)
-40
-20
0
20
40x -40
-20
0
20
40
y
-0.94
-0.92
-0.9
-0.88
-0.86
-0.84
-0.82
-0.8
-0.78
similarity
-0.65
-0.6
-0.55
-0.5
-0.45
-0.4
-0.35
-0.3
-0.25
-3 -2 -1 0 1 2 3
si
m
ila
rit
y
angle
S74161_mattes_large_angle.map
(start 3.1416, end -3.1416, steps 30)
S74161_mattes_large.map
(start -50.0000, end 50.0000, steps 30)
-40
-20
0
20
40x -40
-20
0
20
40
y
-0.65
-0.6
-0.55
-0.5
-0.45
-0.4
-0.35
-0.3
-0.25
similarity
Fig. 4. Optimisation surfaces generated by the normalised correlation metric and by mutual in-
formation on a pair of manually aligned sections from series S74161. The two surfaces on the
left show similarity over rotation angle and over translation for normalised correlation. The two
on the right show similarity over rotation angle and over translation for mutual information.
We plotted a joint histogram for several pairs of sections stained with alternating
dye combinations, as described for section pairs stained with a single dye combination.
The joint histogram in the middle of figure 3 shows an approximately linear (r2 =
0.817) relation, while the relation shown by the histogram on the right seems non-
linear (r2 = 0.740). Although the normalised correlation metric is designed to detect
linear intensity relations, it still shows a distinct minimum and a large capture range for
these manually aligned section pairs. Therefore, we also used the normalised correlation
metric to register section pairs stained with alternating dye combinations. In this case,
scalar conversion contributes to the colour invariance of our registration approach.
Optimisation Method. The papers we reviewed mostly featured optimisation algo-
rithms that do not explicitly compute the partial derivatives of the similarity metric
with respect to the independent variables. Specifically, the downhill simplex method
by Nelder and Mead and Powell’s direction set method seem to be popular. (See [21,
pp.408–420] for a description of these algorithms.)
However, it is possible to compute an approximation of the partial derivatives of
the normalised correlation metric and the mutual information metric. Algorithms that
make use of derivative information often converge in fewer iterations. This does not
necessarily mean that they are always faster, because more computation is needed per
iteration [21, p.395].
We used a popular adaptive step size gradient descent optimiser. During each itera-
tion, this optimiser takes a step in the direction opposite to that of the local gradient. If
the angle between the local gradient at the new position and the local gradient at the old
position is larger than 90 degrees, the optimiser halves the step size.
6 Results
We have reconstructed almost all volumes from the orbita collection. For the purpose of
this section, we focus on two series that will be referred to as S73345 and S74161. The
former consists of 250 sections and the latter of 97 sections. For some of the measure-
ments, we worked with a 100-slice subvolume of S73345. Benchmarks were performed
on a pentium 4, 2.66 GHz, with 1 GB of ram.
In an attempt to speed up the reconstruction we used the centres of gravity of the
binary segmentation masks to estimate an initial registration transformation. The initial

Digitisation and 3D Reconstruction of 30 Year Old Microscopic Sections 645
Fig. 5. Cross-sectional view of original and registered microscopic sections at top and bottom
respectively. Qualitatively, the axial contours of the original 3-D objects seem to have been re-
covered.
Table 2. An overview of the number of iterations and total time required to perform reconstruction
of various series, with and without using the binary segmentation masks to estimate an initial
transform
Series No. of sections Masks No. of iterations Time
S73345-subvol 100 NO 15935 190 min
S73345-subvol 100 YES 3091 23 min
S73345 250 YES 10190 73 min
S74161 97 NO 26350 187 min
S74161 97 YES 2779 22 min
transformation translates the centre of gravity of the floating image onto the centre of
gravity of the reference image. Table 2 shows the performance improvement resulting
from this modification. In the two documented cases we observed speed-ups of a factor
of 5 and 9 respectively. The centre of gravity calculation is rather sensitive to outliers.
Fig. 6. Direct volume rendering of the head of a human foetus created from a reconstructed
dataset. The section-like variations visible are due to small differences in the image intensity
that show up due to the specifics of the transfer function.

646 J.E. van Zwieten et al.
However, the object of this series (the head of a human foetus) has a very regular shape
and therefore an accurate initial transformation could be obtained in most situations.
Figure 5 shows a virtual section along the sagittal plane, perpendicular to the plane
of sectioning, before and after reconstruction. Figure 6 is a direct volume rendering of
the same reconstruction.
7 Conclusions and Future Work
In this paper, we presented a practical procedure for the 3-D reconstruction of a large
set of 30 year old microscopic sections. This procedure consists of a pre-processing
and a registration phase. In the pre-processing phase, the data is resampled, segmented,
and converted to scalar data. We investigated the application of principal component
analysis (PCA) for this conversion. Converting to scalar data allowed the use of standard
similarity metrics in the registration phase. In addition, it facilitated the registration of
consecutive sections stained with alternating dye combinations.
As part of the registration phase, we discussed the selection of a transformation fam-
ily based on the characteristics of the data. We also investigated the relation between
adjacent sections stained with the same and with different dyes.
Finally, we showed a multi-planar view and a direct volume rendering of a recon-
structed foetal head to illustrate the results of our procedure. Visual inspection of the
reconstructions indicates a high overall quality.
An important conclusion is that nitrocellulose embedding, although it is an extremely
labour-intensive process, yields sections that show only minor deformation even after
a period of three decades. As a consequence, rigid registration of this kind of data will
generally yield high quality reconstructions.
The entire collection, including the raw digitised sections and our reconstructions,
will be made available to the public via the Internet. At the moment, we are developing
a web-site and client-server applications for the querying and visualisation of the orbita
collection.
References
1. Ham, A., Leeson, T.: Histology. 5th edn. Pitman medical publishing (1965)
2. Huijsmans, D., Lamers, W., Los, J., Strackee, J.: Toward computerized morphometric facili-
ties: a review of 58 software packages for computer-aided three-dimensional reconstruction,
quantification, and picture generation from parallel serial sections. The Anatomical Record
216 (1986) 449–470
3. van den Elsen, P., Pol, E.J., Viergever, M.: Medical image matching - a review with classifi-
cation. IEEE Engineering in Medicine and Biology (1993) 26–39
4. Maintz, J., Viergever, M.: A survey of medical image registration. Medical Image Analysis
2(1) (1998) 1–36
5. Laan, A., Lamers, W., Huijsmans, D., te Kortschot, A., Smith, J., Strackee, J., Los, J.:
Deformation-corrected computer-aided three-dimensional reconstruction of immunohisto-
chemically stained organs: application to the rat heart during early organogenesis. The
Anatomical Record 224 (1989) 443–457

Digitisation and 3D Reconstruction of 30 Year Old Microscopic Sections 647
6. Verbeek, F.: Three-dimensional reconstruction of biological objects from serial sections
including deformation correction. PhD thesis, Delft Technical University (1995)
7. Kim, B., Boes, J., Frey, K., Meyer, C.: Mutual information for automated unwarping of rat
brain autoradiographs. Neuroimage 5 (1997) 31–40
8. Rydmark, M., Jansson, T., Berthold, C.H., Gustavsson, T.: Computer-assisted realignment
of light micrograph images from consecutive section series of cat celebral cortex. Journal of
Microscopy 165 (1992) 29–47
9. Ourselin, S., Roche, A., Subsol, G., Pennec, X., Ayache, N.: Reconstructing a 3d structure
from serial histological sections. Image and vision computing 19 (2000) 25–31
10. Pitiot, A., Malandain, G., Bardinet, E., Thompson, P.: Piecewise affine registration of biolog-
ical images. In Gee, J., Maintz, J.A., Vannier, M.W., eds.: Second International Workshop on
Biomedical Image Registration (WBIR2003). Volume 2717 of Lecture Notes in Computer
Science., Springer-Verlag (2003) 91–101
11. Guest, E., Baldock, R.: Automatic reconstruction of serial sections using the finite element
method. Bioimaging 3 (1995)
12. Krinidis, S., Nikou, C., Pitas, I.: A global energy function for the alignment of serially
acquired slices. IEEE Transactions on Information Technology in Biomedicine 7(2) (2003)
108–113
13. Galigher, A., Korloff, E.: Essentials of practical microtechnique. Lea and Febiger (1964)
14. Poynton, C.: Color faq – frequently asked questions about color (2002)
15. Young, I., Gerbrands, J., van Vliet, L.: Fundamentals of Image Processing. Delft University
of Technology (1998)
16. Mardia, K.: Statistics of directional data. Probability and Mathemetical Statistics. Academic
Press (1972)
17. Koornneef, L.: Spatial aspects of orbital musculo-fibrous tissue in man. PhD thesis, Amster-
dam University (1976)
18. de Haan, A.: De prenatale ontwikkeling van de humane orbita. PhD thesis (1983-1986)
Unpublished.
19. Roche, A., Malandain, G., Ayache, N.: Unifying maximum likelihood approaches in medical
image registration. International Journal of Imaging Systems and Technology: Special Issue
on 3D Imaging 11 (2000) 71–80
20. Iba´n˜ez, L., Schroeder, W., Ng, L., Cates, J.: The ITK software guide. 2nd edn. (2005)
21. Press, W., Teukolsky, S., Vetterling, W., Flannery, B.: Numerical recipes in C, the art of
scientific computing. 2nd edn. Cambridge University Press (1994)