See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/12468096

Fractal properties of bone

Article in Dentomaxillofacial Radiology · June 2000

DOI: 10.1038/sj/dmfr/4600524 · Source: PubMed

CITATIONS READS

121 814

2 authors, including:

P.F. van der Stelt

Academisch Centrum Tandheelkunde Amsterdam
304 PUBLICATIONS 8,256 CITATIONS

SEE PROFILE

All content following this page was uploaded by P.F. van der Stelt on 19 May 2014.

The user has requested enhancement of the downloaded file.

 Dentomaxillofacial Radiology (2000) 29, 144 ± 153
ã 2000 Macmillan Publishers Ltd. All rights reserved 0250 ± 832X/00 $15.00
www.nature.com/dmfr

REVIEW
Fractal properties of bone
WGM Geraets*,1 and PF van der Stelt1
1
Department of Radiology, ACTA, Academic Centre for Dentistry, Amsterdam, The Netherlands

Objective: To introduce the applications of fractal analysis of bone in medical imaging in

general and dental radiographs in particular.
Results: Various methods for measuring fractal dimension have been used to compare normal
with osteoporotic bone with contradictory results. This disparity may be attributed to
di€erences in the anatomical sites studied and di€erences in methods used to obtain the 2D
images. However, di€erences in methods used to measure fractal dimension can also be
responsible.
Conclusion: When fractal dimension is used to study bone it must be remembered that all
stages in the analytical chain have an impact on the results. Thus, to obtain more conclusive
results, studies on the fractal dimension of bone should be carefully designed and individual
methods thoroughly evaluated.

Keywords: bone and bones; image interpretation, computer-assisted; fractals; osteoporosis

Part 1: Introduction mathematical research with natural sciences and

computing science. Complex images can be created
A century of history by means of simple rules combined with fractal
Although many people consider Benoit Mandelbrot as geometry. This implies that fractal geometry can be a
the father of fractals, a number of predecessors have powerful descriptor of textures in images. The inverse
been identi®ed.1,2 It started in 1875, when the step, of determining the fractal dimension of a given
mathematician du Bois Reymond described continu- texture, is much more complex. In the eighties, the
ous non-di€erentiable functions. This subject was then concepts of fractal geometry were taken up by many
explored further by Cantor, Peano, Koch and Perrin. branches of science and entered dental radiography
In 1900 the statistician Louis Bachelier introduced the from an interest in image pattern recognition and
concept of the random walk or Brownian movement in medical imaging.
his thesis on stock exchange transactions. Perrin's
subsequent work on Brownian motion won him the
Nobel Prize and contributed to the development of Self-similarity and power laws
probability theory. In 1919, Hausdorf and Besicovitch The principle of self-similarity is often explained by
started working on related topics which culminated in citing Richardson who measured the length of the
the publication by Richardson in 1961 on the fractal British coastline by means of a caliper and a set of
behaviour of boundary lines. maps of di€erent scales. He found that the length L
Mandelbrot1 in 1977 brought the concept of increased if he used more detailed maps. Surpris-
`fractals' to the attention of a general audience. He ingly, for most coastlines, a constant D can be
identi®ed families of shapes comprising curves, found between 1 and 2, (and some constant k),
surfaces, disconnected `dusts' and odd shapes. He with L(S)=k. S(17 D) as described in3 ± 5 and many
coined the term `fractal' from the Latine adjective others.
`fractus', meaning `broken'. Mandelbrot's famous As another example, imagine the surface of the
book1 is artistic, rather than scienti®c, connecting moon covered with craters of various sizes. Small
craters are more abundant than big ones. The number
of craters N with an area A or more is found to be
proportional to A71. Photographs of the lunar surface
*Correspondence to: WGM Geraets, ACTA Radiology Department, 2.01, taken at an altitude of one kilometer can hardly be
Louwesweg 1, 1066 EA Amsterdam, The Netherlands discriminated from those taken from one hundred
Received 4 October 1999; accepted 7 February 2000 meters or even one meter. This quality of lunar images
 Fractal properties of bone
WGM Geraets and PF van der Stelt
145
illustrates the essence of self-similarity. By using the of fractal geometry, methods of measurement and
statistical distribution of crater sizes, very realistic applications in microscopy noted the duplication and
landscapes of the moon can be generated, con®rming confusion in the nomenclature of di€erent fractal
that the landscape on the moon to a large extent is dimensions.
ruled by a simple power law between two variables.6
The variety of fractal constructs which exist in
nature, or can be generated by application of the Part 2: Simpli®ed spatial methods
geometric principle of self-similarity, is very great.
Some of the more familiar examples are named after Segmented images
those who studied them in detail: Menger sponges, In many cases, images of bone are partitioned
Cantor dusts, Korcak islands and Lindenmayer (segmented) into bright and dark areas, also referred
systems.2 In general, the most essential fractal to as foreground and background. When viewing
property of self-similarity is associated with one or sections of trabecular bone, the section itself may
more power law relations between the variables already appear to be segmented into foreground and
involved. background. The borderline between foreground and
background in segmented images can be considered as
a curve whose fractal dimension can be measured. If
Fractal dimensions representing roughness the fractal dimension of the original multilevel image is
Pentland7 described an experiment in which eight D and if the conditions of isotropy and uniformity are
images of naturally occurring textures from the ful®lled (which images of bone generally do not meet)
Brodatz collection were shown to 15 subjects who then the fractal dimension of the intersecting boundary
were asked to rank the roughness of the surface. The will be D 71. Within these limitations, several
results were found to agree with the fractal dimension. simpli®ed methods are available for analysing the
Russ2 con®rmed that the fractal dimension summarizes fractal dimension of segmented images.
much of human experience with real, irregular
boundaries and surfaces and that it may be under-
stood as a measure of roughness. However, the visual Caliper Method
perception of roughness is also in¯uenced by the The Caliper Method is probably the best known
contrast of the image.2,5 Changing the contrast of the method to measure the fractal dimension of curves. If
image a€ects its visual appearance but leaves the the length L of a fractal curve is measured with a
fractal dimension the same. However, this ambiguity caliper with span S then: L (S)=k . S(17D). Therefore,
can be avoided by rescaling the range of intensities in when a log-log plot is made of L against S, a straight
the image by dividing them by the standard deviation. line is obtained. If b is the slope of the straight line,
Mandelbrot warned against oversimpli®cation by then the fractal dimension is: D=17b with 15D52
presenting several Cantor dusts, all of which had a and 715b50.2,6,13
fractal dimension of 0.5, but with clearly di€ering Shrout et al.14 used the Caliper Method on dental
roughness.8 Mandelbrot8 proposed the introduction of radiographs of six dry hemimandibles. Although not
another feature which he called lacunarity to mentioned explicitly, it must be assumed that the
encompass this. radiographs were segmented. Regions of interest were
constructed and signi®cant di€erences found between
the specimens. The e€ects of variations in exposure
Variety of methods time, beam angulation and repositioning the region of
In this article we will use the symbol D to refer to the interest were somewhat ambiguous.
fractal dimension. In literature D is referred to as the
Hausdorf dimension9,10 or as the Hausdorf-Besicovich
dimension.11A variety of methods for estimating the Tile Counting Method
fractal dimension has been described in the literature. An expansion of the Caliper Method is the Tile
These methods all measure some characteristic of the Counting Method or Mosaic Amalgation Method.
image (mostly length, surface or volume) as a function Some authors15,16 refer to the Tile Counting Method
of a scale. If the image characteristic is related to the as the Box Counting Method, although the latter is an
scale parameter by a power law, then a log-log plot will expansion of the former. In the Tile Counting Method,
be a straight line with its slope b directly related to D, a succession of square grids of increasing edge length is
as shown in parts 2 ± 4. superimposed on the image to be analysed. Analogous
Another well-known dimension is symbolized as H. with the Caliper Method, it can be shown that the
Dimensions D and H are closely related as D=T+17 number N of tiles contacting the boundary between
H, with T indicating the topological dimension of the foreground and background is related to the edge
situation (T=1,2,3). Considering 05H51, it can be length S by the power law: N (S)=k. S(17D). Therefore,
seen that T5D5T+1. Confusingly, in the literature, when log (N (S)) is plotted against log (S), a straight
both D and H are referred to as the Hausdor€ and line is obtained with slope b. Then D=17b with
Hausdor€-Besicovich dimensions. Cross12 in a review 15D52 and 715b50.6,13

Dentomaxillofacial Radiology
 Fractal properties of bone
WGM Geraets and PF van der Stelt
146
In the Modi®ed Tile Counting Method the total topological dimension, the image becomes a collection
number M of tiles covering the foreground is analysed. of adjacent `skyscrapers'. Methods where the fractal
It can be shown that M is related to edge length S as: dimension is to be measured on a three-dimensional
M (S)=k . S7D. Therefore, when log (M (S)) is plotted surface in space are said to be operating in the spatial
against log (S), a straight line is obtained with slope b. domain.
Then D=7b with 15D52 and 725b571.2,6,15,17,18 A true fractal surface is characterized by indepen-
Van der Stelt and Geraets19,20 used the Tile Counting dence of scale: on every scale the structure is similar. In
Method on segmented radiographs of the radius of an image intensity surface this independency must
perimenopausal women. Majumdar et al.15 used the apply not only to the two spatial dimensions but also
Modi®ed Tile Counting Method on segmented photo- to the intensity dimension.8,24 Whereas the self-similar
micrographs of transiliac crest biopsies and on shapes repeat (either statistically or exactly) under
segmented tomograms of dried human vertebrae. magni®cation, the self-ane shapes repeat only when
They reported a strong positive correlation between coordinates are scaled by di€erent amounts. It is
fractal dimension and trabecular bone area. The same important to note that associating self-similarity with a
methods were used in a follow-up study on transiliac self-ane fractal is implicitly a scaling between
bone specimens.18 Majumdar et al.21 also used the otherwise independent coordinates.6
Modi®ed Tile Counting Method to obtain the fractal
dimension of segmented magnetic resonance images of
the distal radius of perimenopausal women. Haidekker Box Counting Method
et al.16 used the Tile Counting Method on segmented The Box Counting Method relies on progressively
radiographs of vertebrae to study the in¯uence of the coarsening the image. The original image has the ®nest
segmentation threshold value on the fractal dimension. scale approximation and larger scale approximations
with less resolution are used to construct a log-log plot
of surface area against window size.2 Calculate the
Pixel Dilation Method surface area A(1) as the sum of the `roofs' and the
The Pixel Dilation Method starts with converting the exposed sides of the skyscrapers. Then smooth the
foreground to an outline with a width of one pixel. image with a 262 window and calculate the surface
Then the foreground is progressively coarsened by area A(2). Repeat the process with a larger window.
repeated dilation with a circular structuring element of The surface area A can be measured as a function of
width W. Between two dilation steps the length L(W) window size S. It can be shown that A (S)=k . S(27D).
of the outline is estimated as the area of the foreground Therefore, a straight line with slope b can be obtained
divided by W. It can be shown that L(W)=k. W(17D). from a log-log plot of A against S. Then D=27b with
A log-log plot of L against W generates a straight line, 25D53 and 715b50.25
with a slope b. Then D=17b with 15D52 and In the Modi®ed Box Counting Method the number
715b50.2,12,13 N of cubes with side length S needed to cover a fractal
In the Modi®ed Pixel Dilation Method, N(E) is surface is counted. It can be shown that N(S)=k.S7D.
de®ned as the ratio of foreground area to the Therefore a log-log plot of N against S will
structuring element area. It can be shown that approximate a straight line with slope b. Then D=
N(E)=k. E7D/2. A log-log plot of N against E 7 b with 25D53 and 715b50.26,27
generates a straight line, with slope b. Then D=72b Caldwell et al.25 used the Box Counting Method to
with 15D52 and 715b570.5.22,23 study the patterns of the parenchyma in mammograms.
Smith et al.13 studied segmented images of neuronal Sarkar and Chaudhuri28 compared the speed and
cells to compare the Caliper, the Tile Counting and the accuracy of three versions of the Modi®ed Box
Pixel Dilation Method. They found that log-log plots Counting Method (as well as the Blanket Method
of the former were usually curved, whereas the two and the Power Spectrum Method described in section
latter methods gave a reasonable linear relationship. 4), by applying them to textures from the Brodatz
Honda et al.22,23 used the Modi®ed Pixel Dilation collection and synthesized images. Caligiuri et al.29 used
Method to compare segmented normal parotid glands, the Box Counting Method in an attempt to
with those of patients with SjoÈgrens syndrome. They discriminate healthy from osteoporotic lateral lumbar
found D=1.64+0.06 for the former compared with spine radiographs.
D=1.39+0.10 for the latter.

Intensity Variance Method

Part 3: General spatial methods Pentland7 developed the Intensity Variance Method to
test if an image can be approximated by a three-
Image surface dimensional fractal Brownian surface. Later the
Consider the image as a three-dimensional surface by Intensity Variance Method, also known as the
letting the intensity I(x,y) represent the elevation of the `Pentland Method', was used to calculate the fractal
surface above the (x,y) coordinates. Interpreting the dimension. Pentland7 studied the variance (or standard
optical density at point (x,y) as a distance in the third deviation) V in intensity of a fractal surface over
Dentomaxillofacial Radiology
 Fractal properties of bone
WGM Geraets and PF van der Stelt
147
distance S. It can be shown that V(S) is proportional to to classify ultrasonograms of the liver as normal,
S2H. Substituting H=3 7D yields: V(S)=k . S6 ± 2D. hepatoma or cirrhosis. Dubuisson and Dubes27
Now for every distance S calculate the intensity compared the performance of three versions of the
di€erences between any two points at a distance S Blanket Method (described below), two of the Box
from each other and calculate the variance (or Counting Method and the Variation Method with
standard deviation). Then plot V as a function of S square structuring elements. Because the Variation
and ®t a straight line with slope b. Then D=3+b/2 Method performed in the most stable fashion, they
with 25D53 and 725b50.4,7,30,31 used it to segment images from the Brodatz
GaÊrding5 used the Intensity Variance Method to collection. Huang et al.26 compared the accuracy
segment test images. Veenland et al.24,31,32 examined the and robustness of the modi®ed Box Counting
e€ect of noise and blur on the outcome of the Intensity Method and the Variation Method, using square
Variance Method (as well as the Variation Method, the structuring elements but found no de®nite advantage
Blanket Method and the Power Spectrum Method (see for either method.
part 4). They found that the choice of screen ®lm
combination can result in a considerable underestima-
tion of 0.8 in the fractal dimension. Blanket Method
The Blanket Method is considered a modi®cation of
the Variation Method. It was ®rst described by Peleg et
Hurst Method al.3 In order to make the calculations robust, the area
To construct a Hurst plot, the data are scanned to ®nd of the image intensity surface is calculated by means of
the two points within a distance S whose di€erence M two blankets. The upper blanket represents the
is greatest. For purpose of normalization M is divided brightest value in the neighborhood of any pixel,
by the standard deviation of the elevation data. It can whereas the lower blanket represents the darkest. The
be shown that M(S)=k. S(37D) The slope b of the volume V enclosed by the blankets depends on the size
straight line obtained by plotting M as a function of S L of the neighborhood, and the area A is de®ned as
is used to calculate: D=37b with 25D53 and (V(L)7V(L71))/2. It can be shown that A
05b51.2,10 (L)=L(27D). A straight line with slope b is obtained
from a log-log plot of A against L. Then calculate:
D=27b with 25D53 and 715b50.3,31
Variation Method Peleg et al.3 used the Blanket Method, with circular
In the Variation Method, the maximum di€erence in structuring elements, to study textures from the
intensity within the neighborhood of a pixel is related Brodatz collection. In addition rod-shaped structuring
to the size of the neighborhood. If the shape of the elements were used as described below. Kaufman et
neighborhood is circular, then the shape is expressed al.34 applied the Blanket Method to radiographs of a
by the radius; in the case of square neighborhoods, the healing fracture of the calcaneus. Although not
edge length is used. If the shape of the chosen explicitly stated, they appear to have used round
structuring element (neighborhood) is circular, then D structuring elements. Lynch et al.33 modi®ed Peleg's
is called the Minkowski dimension.2 Sometimes the Blanket Method by incorporating dilation and erosion
vertical scale is normalized before measuring D, as in operations and elliptical structuring elements. They
the Hurst Method described above. The total variation reported a fractal dimension of 2.55 from macroradio-
V is de®ned as the sum of all maximum di€erences, graphs of healthy knees; in osteoarthritis this value
determined for each pixel in its neighborhood. If L increased to 2.69 in parallel with the severity of the
indicates the size of the neighborhood it can be shown: disease. Wilding et al.35 used the Blanket Method with
V(L)=k. L(37D). A log-log plot of V against L a square structuring element to examine the e€ect of
generates a straight line with slope b. Then D=37b implant placement on alveolar bone. A signi®cant
with 25D53 and 05b51.26,27 increase in fractal dimension especially around the neck
The calculations used in the Variation Method are of the implant was found for up to 2 years after
also performed in the Morphological Method by implantation.
dilation and erosion operators which were originally
de®ned for binary images and now generalized as
minimum and maximum operators. These operators Part 4: Spectral methods
examine the neighborhood of each pixel in the original
image, ®nd the brightest and darkest values in the Spatial and spectral methods
neighborhood and create a new image in which the The methods to calculate fractal dimension can be
central pixel is given the di€erence between the two divided into two types: spatial and spectral. The ®rst
extreme values. The brightness values of the di€erence type operates in the spatial domain and includes the
image are integrated and the operation is repeated for methods described above. The second type operates in
the next neighborhood size.33 the frequency domain, using the Fourier power
Wu et al.30 used a modi®ed version of the spectrum. The two types are uni®ed by the principles
Variation Method, with square structuring elements, of fractional Brownian motion.

Dentomaxillofacial Radiology
 Fractal properties of bone
WGM Geraets and PF van der Stelt
148
Power Spectrum Method that fractal dimension alone is insucient to segment
In the Power Spectrum Method, the two-dimensional natural textures.
Fourier transform of the digital image is calculated ®rst Pentland7 used the Power Spectrum Method on
and, then the two-dimensional power spectrum derived. blocks of 868 pixels to segment photographs of
The two-dimensional power spectrum is reduced to a natural scenes. He reported that calculations of the
one-dimensional power spectrum by averaging over local fractal dimension yielded values less than the
circles.11,24,36,37 The power spectrum varies with fre- topological dimension, especially near the boundaries
quency as P(f)=k . f(71 ± 2H) with 05H51.5,7,24,31,36,38,39 of a texture, which according to theory are illegal. He
So, when log (P(f)) is plotted against log (f), a straight suggested that this feature could be used to detect the
line can be ®tted with slope b=(71 ± 2H). Therefore, edge of a texture (see also39). Kuklinski et al.39 used
H=7(1+b)/2. Substitution of H=3 ± D yields: the same method on blocks of 16616 pixels in
D=(7+b)/2 with 25D53 and 735b571. Some radiographs of dry human jaws, trying to identify
authors prefer to use positive values of b. Then bony regions by means of several features including
D=(7 ± b)/2 with 25D53 and +15b5+3.6,11,37,40 the fractal dimension. Although the Power Spectrum
Ruttimann et al.36 used the Power Spectrum Method Method should yield values of D between 2 and 3,
to study radiographs of interdental alveolar bone of 12 they found it was between 3 and 4. Lalitha and
perimenopausal women. Radiographs were also ob- Majumder38 compared the e€ects of di€erent magni-
tained at three projection angles of sections of human ®cation techniques on two remote sensing images. The
mandible before and after the decalci®cation for fractal dimension was calculated on blocks of 868
35 min in 10% formic acid (see also41,42). Samaraban- pixels. The reported values of D between 1 and 2 were
du et al.11 used it on radiographs of rat femurs and lower than expected with the Power Spectrum
Law et al.43 to analyse dental radiographs of Method.
perimenopausal women. They reported mean values
of D=3.19 for healthy women, and D=3.24 for those
with osteoporosis. Both values exceeded the theoretical Directional measurements with spatial methods
maximum of 3.0 which the authors attribute to Bone is a natural material which is designed to resist
measuring errors. Berry et al.40 applied this procedure gravitational and mechanical forces and thus, has an
to radiographs of vertebral bone exposed to 10% nitric internal structure that is highly oriented (anisotropic).
acid. The fractal dimension increased rapidly from 2.0 The anisotropic features of bone structure, can be
to 3.0 for exposures of up to 12 min. expressed by directional measurements of fractal
dimension. The methods described above can be
modi®ed to operate on lines instead of images. When
Part 5: Special applications measured on a line, fractal dimension D should have
values between 1 and 2.
Localized measurements Southard et al.45 used a special linear version of
The methods described above can be applied to the Caliper Method to study radiographs of sections
the complete image to yield a global value of the of vertebra exposed to 0.1 N hydrochloric acid for up
fractal dimension. In principle, local measurements to 90 min. They reported that decalci®cation reduced
of D can be done by isolating several regions of the roughness and the fractal dimension decreased
interest, which then can be analyzed by a global from 1.26 to 1.1. Pentland7 used the Intensity
method. However, there are more sophisticated Variance Method to make local directional measure-
ways for localized applications of the methods ments. For purpose of segmentation he made
described above. Mostly, these methods are used measurements in two directions on blocks of 16616
for texture classi®cation and image segmentation. pixels. He reported that the two-parameter fractal
For segmentation a threshold value must be segmentation had a classi®cation accuracy of 84%.
calculated which most authors derive from the Samarabandu et al.11 used the Morphological Method
histogram of the fractal values. with rod-shaped structuring elements on radiographs
Dellepiane et al.44 used both the Intensity Variance of rat femur to measure the fractal dimension in two
Method and the Blanket Method (with a square directions.
structuring element) on regions of 666 pixels to Peleg et al.3 applied the Blanket Method, using
segment two radiographs of human lung. The two circular as well as rod-shaped structuring elements, to
methods produced similar fractal images with slightly compare textures from the Brodatz collection. The
di€erent segmentation. GaÊrding5 used the Intensity rod-shaped elements were used to make measure-
Variance Method on blocks of 16616 pixels to ments in four directions. They demonstrated varia-
segment test images. Chen et al.10 used the Hurst tions in slope within a short range of scales. Lynch et
Method on regions of 767 pixels to segment al.33 used the same method to compare macroradio-
ultrasonograms of liver and chest radiographs. graphs of healthy with osteoarthritic knees, making
Dubuisson and Dubes27 used the Variation Method measurements in two directions with rod-shaped
on regions of 32632 pixels to segment images from structuring elements. Although this type of direc-
the Brodatz collection. They concurred with GaÊrding5 tional measurement should yield values of D between
Dentomaxillofacial Radiology
 Fractal properties of bone
WGM Geraets and PF van der Stelt
149
1 and 2, they found results of up to 3.6. The same decreased from 0.42 to 0.33 and then after several
Blanket Method was used by Buckland-Wright et months increased to 0.55. Using D=3 ± H, it can be seen
al.46 to study macroradiographs of the lumbar spine that D increased in the ®rst 3 months and then decreased
of perimenopausal women. Using rod-shaped struc- to a lower value than before the injury. Jacquet et al.48
turing elements they made measurements in horizon- performed measurements in 36 directions on radiographs
tal and vertical directions and compared them with of decalci®ed femoral bone and fractured human
bone mineral density and body weight. Their values calcaneus. H decreased from 0.6 to 0.3 after 8 h
for D were between 2 and 3. Wilding et al.35 used decalci®cation. In the ®rst 3 months after injury H
four rod-shaped structuring elements with the Blanket decreased from 0.5 to 0.3. It increased again to 0.6 after 6
Method to obtain measurements in four directions. months of mobility. Using D=2 ± H, it can be seen that
They also obtained values for D of between 2 and 3. D increases initially and then decreases to a value lower
In both cases the results exceeded the theoretical than before the injury. Fortin et al.47 also measured H in
values of between 1 and 2. 36 directions in radiographs of human femur exposed to
acid, and of a fractured human heel. These results were
similar to those of Jacquet et al.48 Benhamou et al.49
Directional measurements with spectral methods measured H as an average of 36 directions in radiographs
In the linear version of the power spectrum method, of the calcaneus. They found coecients of variations of
straight lines are analysed in various directions. The less than for 1% intra- and interobserver precision and
power spectrum of data on a scanning line is related to about 2% when successive radiographs were analysed.
the frequency by the same power law as described
above. When the slope b is determined of the linear
approximation of the power spectrum then b = (71 ± Part 6: Problems in fractal geometrical analysis
2H). When analysing a scanline, the topological
dimension is 2 and D=2 ± H. Combining the two Methods are di€erent
formulas yields D=(5+b)/2 with 15D52 and The methods described above, although related, do not
735b571. Some authors, preferring to use positive all measure the same thing. They are not mathematically
values for b arrive at D=(57b)/2 with 15D52 and identical and give rise to di€erent numerical values. Only
15b53.6,11,37 when they are applied to an ideal fractal surface
Chen et al.37 used a one-dimensional version the (continuous and truly self-similar), would the results be
Power Spectrum Method to perform directional the same. This is not true for digitized pixel data, even
measurements along two directions in three regions of without bias and errors. The e€ect on the reliability of
interest on a radiograph. The radiograph was then measurements of fractal dimension when images are
rotated 908 and the analysis was repeated. Surprisingly, sampled rather than continuous is described in4,5. In most
the second series of measurements was found to deviate cases, it is only possible to compare numerical values
from the ®rst. The authors suggest this e€ect was from di€erent pro®les or surfaces using one type of
caused by the hardware used to digitize the image. measurement.2,24,26,27 Several researchers have compared
the available methods on images of known fractal
dimension varying from 2.1 up to 2.9. Compared with
Directional measurements with Maximum Likelihood the Power Spectrum Methods, all the other methods and
Method especially the Variation Method, result in a reduced
Lundahl et al.4 and Fortin et al.47 extended with great range of estimates of the fractal dimension. Nevertheless,
care the mathematical model of Fractional Brownian the estimates vary monotonically with fractal dimen-
Motion from continuous to discrete data. They started sion13,24,26 ± 28, except for a Box Counting Method in27,
from the general models of Fractional Brownian yielding a very narrow range (less than 0.03) with a
Motion and Fractional Gaussian Noise to derive an negative correlation with the known fractal dimension.
estimation of H from a scanline in an image. In the
continuous case the power spectrum of Fractional
Gaussian Noise varies with frequency as P(f)=k . Fitting a straight line
f(+1 ± 2H) with 05H51. From this, the most reliable There is a problem in ®tting a straight line in the log-log
approximation of H in the discrete case is derived.4 plot in order to obtain its slope. True fractal objects have
They proposed using the Intensity Variance Method as the property of self-similarity over all possible ranges of
an alternative to the Maximum Likelihood Method, in scales, whereas self-similarity in bone is con®ned to sizes
order to avoid high computational complexity at the between about 0.1 mm and 5 cm.
price of lower reliability.4 Real images and surfaces can not be true
Lundahl et al.4 measured the fractal dimension in four mathematical fractals because the latter are de®ned to
directions in heeling fractures of the calcaneus. They exist at all scales. Physical surfaces, in contrast, have
localized the measurements by breaking up the images an overall size that places an upper limit on the range
into separate blocks. Composite values of H were of applicable scales. A lower limit is set by the size of
obtained by averaging the values obtained in each the particles that make up the surface. We can
direction over each block. They found H slowly therefore say that a surface is a fractal if the fractal

Dentomaxillofacial Radiology
 Fractal properties of bone
WGM Geraets and PF van der Stelt
150
dimension is stable over a wide range of scales, the bone mass. However, the e€ects of osteoporosis on
implication being that it can be accurately approxi- bone structure are less well established. Osteoporotic
mated over that range of scales by a single fractal dissolution of bone structure may consist of several
function.7 changes. Thickness and porosity of the cortex can alter,
When the log-log plot of the image consists of two and trabeculae can be thinned, perforated or lost
(or more) parts that can be approximated by two (or completely. According to Buckland-Wright et al.46
more) straight line segments it is called a multi-fractal. osteoporosis of vertebrae starts with focal perforation
Examples of multifractal behavior are provided of coarse vertical trabecular plates and progresses by
by29,50,51. converting these fenestrated plates into a lattice of bars
and rods. In the more advanced stages the horizontal
struts in the spine disappear. On radiographs the
Noise sensitivity vertical trabeculae become more accentuated. The
A problem in fractal analysis of binary images is the trabecular pattern in radiographs of the hips can be
e€ect of noise. This is discussed in2,4,12,24,28,31,32 and graded to yield the Singh index52 which correlates to
illustrated in Figure 1. some extent with the degree of osteoporosis. A number
of workers have attempted using acid decali®cation as
a model for osteoporosis.30,40,45,47,48,53
Part 7: Osteoporosis

The e€ects of osteoporosis on the bone mass and Osteoporosis associated with low D
structure have been the subject of many studies. It is Southard et al.45 found that the average fractal
well known that osteoporosis is associated with a low dimension decreased from 1.26 to 1.10 in radiographs

Figure 1 A 10610 mm detail from a radiograph of the distal radius to which di€erent amounts of low and high frequency noise have been
added. The e€ect on the fractal dimensions is shown in the upper left-hand corner

Dentomaxillofacial Radiology
 Fractal properties of bone
WGM Geraets and PF van der Stelt
151
of decalci®ed human alveolar bone. Lee et al.53 postmenopausal women with a matched control
measured fractal dimension on linear and rectangular group yielding Dcontrol=1.261 and Dosteoporotic=1.309.
regions of interest on radiographs of decalci®ed cross- Van der Stelt and Geraets19 studied the radiographic
sections of bovine femur. D in the linear region trabecular pattern of the distal radius of 294
decreased during decalci®cation whereas the rectangu- perimenopausal women at annual intervals over a
lar region failed to show any trend. Majumdar et al.15 period of 3 years. They reported that there was a
compared segmented photomicrographs of transiliac weak positive correlation between fractal dimension
crest biopsies of three normal with three osteoporotic and time.
subjects. The average fractal dimension in the normal
subjects was 1.62 compared with 1.20 in the
osteoporotic patients. Weinstein and Majumdar18 in a Discussion
follow up-study of15 compared 12 subjects in each
group Dcontrol was=1.329 compared with Dosteoporotic of Although acid-induced decalci®cation of bone may be a
1.224. Majumdar et al.21 determined the fractal poor model for osteoporosis, it has the great advantage
dimension from MRI of the distal radius of three of ease of access, enabling the study of structural changes
groups of women: 10 premenopausal, nine non-fracture over time. The discrepancy between experiments
postmenopausal and 11 osteoporotic postmenopausal. reporting low D45,53 and high D36,40,48 may originate
They found that the osteoporotic group had the widest from di€erences in the acid used, the speed of
trabecular spacing and the lowest fractal dimension: decalci®cation and other experimental conditions
Dpre=1.49, Dnon-fracture=1.43, and Dosteoporotic=1.34. determining the fenestration of trabecular plates and
Fazzalari and Parkinson51 analysed thin sections of the loss of entire trabecular branches. Southard et al.45
the iliac crest of 64 postmenopausal women, 31 with suggest that anatomical di€erences may account for the
vertebral crush fracture and 33 without. The Box contradictory results between the maxilla45 and mand-
Counting Method yielded a multifractal behavior ible.36 They argue that `pitting' may have occurred in
consisting of three parts of which only the central thicker mandibular cortex but not in maxillar where it is
part discriminated between the nonfracture and thinner. The fact that in the experiments of Ruttiman et
fracture groups: Dnonfracture=1.23, Dfracture=1.15. al.36 17% of the samples had lowered D is another
indication of the in¯uence of anatomical di€erences.
Another possible explanation for the discrepancies in
Osteoporosis associated with high D the experiments relating bone structure to osteoporosis
Berry et al.40 decalci®ed dried human vertebral bone. may lie in the variety of imaging modalities used.
Digitized radiographs were made every 30 s and the Comparison of radiographs with thin sections of bone
fractal dimension increased from 2.0 to 3.0. Rutti- or transverse sections obtained by MR is complicated
mann et al.36 found an increase of the fractal by the anisotropic nature of bone. Each imaging
dimension in 83% of their samples during decalcifica- modality may have its own nonlinear artefacts such
tion. This was con®rmed by an in vivo comparison of as sampling frequency and noise. It is striking that
intra-oral radiographs of six premenopausal with six most reports based on radiographs in vivo ®nd an
postmenopausal women: Dpre=2.68 whereas association between osteoporosis and increased values
Dpost=2.82. Jacquet et al.48 reported that H of of D. This pattern is found with both di€erent bones
decalci®ed femoral head decreased from 0.6 to 0.3 in (calcaneus,4,34,54 mandible,36,43 and radius19) and di€erent
8 h. After fracture of the calcaneus H decreased from methods to measure fractal dimension (Power Spec-
0.5 to 0.3 during 3 months in plaster, and then 6 trum Method,36,43 Maximum Likelihood Method,4,54
months later it had increased to 0.6. Using D= 3 ± H, Tile Counting Method,19 Box Counting Method34).
it can be seen that immobility leads to an increase in Southard et al.45 also suggest that the di€erence
D (see also47). Law et al.43 compared dental radio- between their results and those of Ruttiman et al.36
graphs of the mandibles of osteoporotic with normal may be caused by di€erences in methods of measuring
control subjects (n=60, 44 respectively). They fractal dimension; they used the caliper method on
reported higher fractal dimensions for the osteoporo- vertical scan lines whereas the latter used power
tic group: Dcontrol=3.19, Dosteoporotic=3.24. Lundahl et spectrum analysis of a square region of interest. The
al.4 investigated one patient with a healing fracture of experiments of Lee et al.53 suggest that the shape of
the calcaneus. During 3 months of immobilization, H the region of interest can have a profound e€ect.
decreased from 0.42 to 0.33 (D increasing from 2.58 Majumder et al.50 radiographed cubes of trabecular
to 2.67), whereas 6 months after removal of cast, H bone from the spine from three orthogonal directions
had increased to 0.55 (D decreasing to 2.45). These and measured fractal dimensions by the Box Counting
results are consistent with Jacquet et al.48 Similarly Method and the Intensity Variance Method. The
Kaufman et al.34 reported D increased from 2.51 to measurements were found to have a signi®cant
2.55 during the initial stages of disuse osteoporosis of negative correlation along two of three directions
the calcaneus. Two years later when the fracture had and were uncorrelated for the third direction. These
healed, D had decreased to 2.49. Pothuaud et al.54 results can possibly be explained by the narrow range
compared radiographs of the calcaneus of 39 of the measurement values.

Dentomaxillofacial Radiology
 Fractal properties of bone
WGM Geraets and PF van der Stelt
152
In conclusion, the diversity of the results described in sion of bone, future studies should be carefully
section 7 may be explained by any or all of anatomical designed and their methods thoroughly tested.
di€erences, di€erences in the methods used to obtain
2D images of 3D bone and di€erences in the Acknowledgements
techniques for measuring fractal dimension. In order We thank Dr Rachid Jennane for his helpful comments on
to obtain unambiguous conclusions on fractal dimen- the theory of fractals.

References

1. Mandelbrot BB. The fractal geometry of nature. 3rd ed. New 21. Majumdar S, Genant HK, Grampp S, Newitt DC, Truong VH,
York: WH Freeman and company, 1983. Lin JC et al. Correlation of trabecular bone structure with age,
2. Russ JC. Fractal Surfaces. 1st ed. New York: Plenum Press, bone mineral density, and osteoporotic status: in vivo studies in
1994. the distal radius using high resolution magnetic resonance
3. Peleg S, Naor J, Hartley R, Avnir D. Multiple resolution texture imaging. J Bone Mineral Res 1997; 12: 111 ± 118.
analysis and classi®cation. IEEE Trans on Pattern Analysis and 22. Honda E, Domon M, Sasaki T. A method for determination of
Machine Intelligence 1984; 6: 518 ± 523. fractal dimensions of sialographic images. Investigative Radi-
4. Lundahl T, Ohley WJ, Kay SM, Si€ert R. Fractional Brownian ology 1991; 26: 894 ± 901.
motion: a maximum likelihood estimator and its application to 23. Honda E, Domon M, Sasaki T, Obayashi N, Ida M. Fractal
image texture. IEEE Trans Med Imag 1986; 5: 152 ± 161. Dimensions of ductal patterns in the parotid glands of normal
5. GaÊrding J. Properties of fractal intensity surfaces. Pattern subjects and patients with SjoÈgren Syndrome. Investigative
Recognition Letters 1988; 8: 319 ± 324. Radiology 1992; 27: 790 ± 795.
6. Barnsley MF, Devaney RL, Mandelbrot BB, Peitgen HO, Saupe 24. Veenland JF, Grashuis JL, van der Meer F, Beckers ALD,
D, Voss RF. The Science of Fractal Images. 1st ed. New York: Gelsema ES. Estimation of fractal dimension in radiographs.
Springer-Verlag, 1988. Medical Physics 1996; 23: 585 ± 594.
7. Pentland AP. Fractal-based description of natural scenes. IEEE 25. Caldwell CB, Stapleton SJ, Holdsworth DW, Jong RA, Weiser
Trans on Pattern Analysis and Machine Intelligence 1984; 6: 661 ± WJ, Cooke G et al. Characterisation of mammographic
674. parenchymal pattern by fractal dimension. Physics in Medicine
8. Mandelbrot BB. A Fractal's lacunarity and how it can be tuned and Biology 1990; 35: 235 ± 247.
and measured. In: Nonnenmacher TF, Losa GA, Weibel ER, 26. Huang Q, Lorch JR, Dubes RC. Can the fractal dimension of
(ed). Fractals in biology and medicine. Basel: BirkhaÈuser Verlag, images be measured? Pattern Recognition 1994; 27: 339 ± 349.
1993, pp 8 ± 21. 27. Dubuisson MP, Dubes RC. Ecacy of fractal features in
9. West BJ, Goldberger AL. Physiology in fractal dimensions. segmenting images of natural textures. Pattern Recognition
American Scientist 1987; 75: 354 ± 365. Letters 1994; 15: 419 ± 431.
10. Chen CC, Daponte JS, Fox MD. Fractal feature analysis and 28. Sarkar N, Chaudhuri BB. An ecient approach to estimate
classi®cation in medical imaging. IEEE Trans Med Imaging fractal dimension of textural images. Pattern Recognition 1992;
1989; 8: 133 ± 142. 25: 1035 ± 1041.
11. Samarabandu J, Acharya R, Hausmann E, Allen K. Analysis of 29. Caligiuri P, Giger ML, Favus MJ. Multifractal radiographic
bone X-rays using morphological fractals. IEEE Trans Med analysis of osteoporosis. Medical Physics 1994; 21: 503 ± 508.
Imaging 1993; 12: 466 ± 470. 30. Wu CM, Chen YC, Hsieh KS. Texture features for classi®cation
12. Cross SS. The application of fractal geometric analysis to of ultrasonic liver images. IEEE Trans Med Imag 1992; 11: 141 ±
microscopic images. Micron 1994; 25: 101 ± 113. 152.
13. Smith TG, Marks WB, Lange GD, Sheri€ WH, Neale EA. A 31. Veenland JF. Texture analysis of the radiographic trabecular
fractal analysis of cell images. J Neurosci Methods 1989; 27: bone pattern in osteoporosis (dissertation). Rotterdam: Erasmus
173 ± 180. University, 1999.
14. Shrout MK, Potter BJ, Hildebolt CF. The e€ect of image 32. Veenland JF, Grashuis JL. Texture analysis in radiographs: the
variations on fractal dimension calculations. Oral Surg Oral in¯uence of modulation transfer function and noise on the
Med Oral Pathol Oral Radiol Endod 1997; 84: 96 ± 100. discriminative ability of texture features. Medical Physics 1998;
15. Majumdar S, Weinstein RS, Prasad RR. Application of fractal 25: 922 ± 936.
geometry techniques to the study of trabecular bone. Medical 33. Lynch JA, Hawkes DJ, Buckland-Wright JC. Analysis of texture
Physics 1993; 20: 1611 ± 1619. in macroradiographs of osteoarthritic knees using the fractal
16. Haidekker MA, Andresen R, Evertsz CJG, Banzer D, Peitgen signature. Phys Med Biol 1991; 36: 709 ± 722.
HO. Assessing the degree of osteoporosis in the axial skeleton 34. Kaufman JJ, Hakim N, Nasser P, Mont M, Klion M, Pilla AA et
using the dependence of the fractal dimension on the grey level al. Application of digital image processing and pattern
threshold. Br J Radiol 1997; 70: 586 ± 593. recognition to diagnosis of osteoporosis. IEEE Proceedings of
17. Chung HW, Chu CC, Underweiser M, Wehrli FW. On the the Annual Conference of the Engineering in Medicine and Biology
fractal nature of trabecular structure. Medical Physics 1994; 21: Society 1987; 9: 1334 ± 1335.
1535 ± 1540. 35. Wilding RJC, Slabbert JCG, Kathree H, Owen CP, Crombie K,
18. Weinstein RS, Majumdar S. Fractal geometry and vertebral Delport P. The use of fractal analysis to reveal remodelling in
compression fractures. J Bone Mineral Res 1994; 9: 1797 ± 1802. human alveolar bone following the placement of dental
19. van der Stelt PF, Geraets WGM. The fractal dimension of the implants. Arch of Oral Biol 1995; 40: 61 ± 72.
trabecular pattern in patients with increased risk of alveolar 36. Ruttimann UE, Webber RL, Hazelrig JB. Fractal dimension
ridge resorption. Proc Ann Int Conference IEEE/EMBS 1990; 12: from radiographs of peridental alveolar bone: a possible
2071 ± 2072. diagnostic indicator of osteoporosis. Oral Surg Oral Med Oral
20. van der Stelt PF, Geraets WGM. Digital quanti®cation of the Pathol 1992; 74: 98 ± 110.
trabecular pattern and the e€ects of magni®cation. J Dent Res
1991; 70: 528 (abstract 2096).

Dentomaxillofacial Radiology
 Fractal properties of bone
WGM Geraets and PF van der Stelt
153
37. Chen J, Zheng B, Chang YH, Shaw CC, Towers TC, Gur D. 46. Buckland-Wright JC, Lynch JA, Rymer J, Fogelman I. Fractal
Fractal analysis of trabecular patterns in projection radio- signature analysis of macroradiographs measures trabecular
graphs. An assessment. Investigative Radiol 1994; 29: 624 ± 629. organization in lumbar vertebrae of postmenopausal women.
38. Lalitha L, Majumder DD. Fractal based criteria to evaluate the Calci®ed Tissue International 1994; 54: 106 ± 112.
performance of digital image magni®cation techniques. Pattern 47. Fortin C, Kumaresan R, Ohley W, Hoefer S. Fractals and
Recognition Letters 1989; 9: 67 ± 75. scaling theory. IEEE Engineering in Medicine and Biology 1992;
39. Kuklinski WS, Chandra K, Ruttiman UE, Webber RL. 11: 65 ± 71.
Application of fractal texture analysis to segmentation of 48. Jacquet G, Ohley WJ, Mont MA, Si€ert R, Schmukler R.
dental radiographs. Proc. SPIE Medical Imaging III: Image Measurement of bone structure by use of fractal dimension.
Processing 1989; 1092: 111 ± 117. Annual International Conference of the IEEE Engineering in
40. Berry JL, Towers JD, Webber RL, Pope TL, Davidai G, Medicine and Biology Society 1990; 12: 1402 ± 1403.
Zimmerman M. Change in trabecular architecture as measured 49. Benhamou CL, Lespessailles E, Jacquet G, Harba R, Jennane R,
by fractal dimension. J Biomechanics 1996; 29: 819 ± 822. Loussot T. Fractal organization of trabecular bone images on
41. Ruttiman UE, Ship JA. The use of fractal geometry to calcaneus radiographs. J Bone Mineral Res 1994; 9: 1909 ± 1918.
quantitate bone structure from radiographs. J Dent Res 1990; 50. Majumdar S, Lin J, Link T, Millard J, Augat P, Ouyang X et al.
69: 287 (abstract 1431). Fractal analysis of radiographs: assesssment of trabecular bone
42. Webber RL, Hazelrig JB, Patel JR, Van den Berg HR, Lemons structure and prediction of elastic modulus and strength.
JE. Evaluation of site-speci®c di€erences in trabecular bone Medical Physics 1999; 26; 1330 ± 1340.
using fractal geometry. J Dent Res 1991; 70: (special issue) 528 51. Fazzalari NL, Parkinson IH. Fractal properties of cancellous
(abstract 2095). bone of the iliac crest in vertebral crush factors. Bone 1998; 23;
43. Law AN, Bollen AM, Chen SK. Detecting osteoporosis using 53 ± 57.
dental radiographs: a comparison of four methods. J Am Dent 52. Singh M, Nagrinth AR, Maini PS. Changes in trabecular
Assoc 1996; 127: 1734 ± 1742. patterns of the upper end of the femur as an index of
44. Dellepiane S, Serpico SB, Vernazza G, Viviani R. Fractal ± osteoporosis. J Bone Jt Surg 1970; 52A: 457 ± 467.
based image analysis in radiological applications. Proc SPIE 53. Lee KI, Choi SC, Park TW, You DS. Fractal dimension
Visual Communications and Image Processing II 1987; 845: 396 ± calculated from two types of region of interest. Dentomaxillofac
403. Radiol 1999; 28; 284 ± 289.
45. Southard TE, Southard KA, Jakobsen JR, Hillis SL, Najim CA. 54. Pothuaud L, Lespessailles E, Harba R, Jennane R, Royant V,
Fractal dimension in radiographic analysis of alveolar process Eynard E et al. Fractal analysis of trabecular bone texture on
bone. Oral Surg Oral Med Oral Pathol Oral Radiol Endod 1996; radiographs: discriminant value in postmenopausal osteoporo-
82: 569 ± 576. sis. Osteoporosis Int 1998; 8; 618 ± 625.

Dentomaxillofacial Radiology

View publication stats