jj&KSj&SjlK Dentomaxillofac. Radiol , Vol. 25, No. 1, pp.

42-47, 19%
Copyright © 1996 Elsevier Science Ltd. for the IADMFR
iRSSiajl^ Printed in Great Britain. All rights reserved
λΓΛΕΛΙ 0250-832X/96 $15.00 + 0.00

ELSEVIER

0250-832X(95)00031-3

Histogram-matching and histogram-

flattening contrast correction methods:
a comparison
C.F. Hildebolt*, R.K. Walkup*, G.L. Conoverf, N. Yokoyama-Crothers*, T.Q. Bartlett*,
M.W. Vannier*, M.K. Shroud and J.J. Camp§
* Washington University, St Louis, Missouri,+ Southern Illinois University, Alton, Illinois, * Medical College of
Georgia, Augusta, Georgia, and § Mayo Foundation, Rochester, Minnesota, USA

Received 30 December 1994 and in final form 4 May 1995

Objectives. To compare the results of two methods of histogram matching and two methods of
histogram flattening for their ability to correct for contrast variations in digital dental images.
Methods. A custom-built, aluminium stepwedge with 0.1,0.5 and 1.0 mm steps was placed
over Ektaspeed films and exposed for 0.06,0.12 and 0.25 s, respectively. Radiographs were
digitized at 50 μπι spatial resolution and 12-bit contrast resolution. Contrast corrections were
performed using Rüttimann et al.'s algorithm (1986) for one method of matching (RM) and
flattening (RF) and Castleman's algorithm (1979) for the other method of matching (CM) and
flattening (CF). Mean pixel grey-scale values were determined for each step. The 0.12 s
exposure was considered to be the target image exposure. Absolute differences in pixel
grey-scale values between the target images and the modified images were determined.
Results. The median values of the absolute differences in pixel grey-scale values between the
target images and the contrast corrected images were: CM = 4.3; RM = 4.1; CF = 70.2 and
RF = 70.2.
Conclusion. Castleman's and Rüttimann's matching algorithms perform equally well in
correcting digital image contrast. Histogram flattening was less effective.

Keywords: radiographic image enhancement; subtraction radiography; signal processing, computer-

assisted

Dentomaxillofac. Radiol., Vol. 25, N o . 1, 42-47, 1996

Retrospective correction for variations in dental film
exposure, film processing and radiograph digitization is
an essential step in measurement of small changes in
hard tissue 1-4 . In 1986, a non-parametric histogram-
matching algorithm for making such corrections was
introduced by Rüttimann et al.5. This method, which
we refer to as Rüttimann's algorithm, has found wide
acceptance in quantitative digital-imaging studies 6-9 .
Additional algorithms have been reviewed, and those
incorporating non-parametric contrast correction, simi-
lar to that described by Rüttimann, have performed
best 10 · 11 .
The first step in using the Rüttimann matching (RM)
algorithm is to create a histogram for each of the digital
images that are to undergo contrast correction. For
each histogram, a running sum of the frequencies of
pixels at each grey-scale value is calculated to yield a
cumulative grey-scale distribution (cumulative density
function; CDF). The cumulative sums of pixels at each
grey-scale value in the input (original) image are
matched to the corresponding cumulative sums of the
target (desired) image. The result is that the cumulative
sum of the desired image is less than or equal to the
cumulative sum of the original image while also being
greater than the next lower bin of the cumulative sum
of the original image 5 .
A similar approach to histogram matching contrast
correction is integrated into the Mayo Clinic's
ANALYZE™ program - a comprehensive software
system of biomedical and scientific visualization
applications 12 ' 13 . The matching routine in this program
is based on an algorithm described by Castleman in
197914. With this algorithm, the CDF of the input
image is forced to match as closely as possible that of
the target image, and, as such, the algorithm is some-
what similar to that of Rüttimann's. Over the last two
Comparison of contrast-correction methods
years, we have used both algorithms to perform con-
trast corrections. In preliminary comparisions, it
appeared that Castleman's algorithm out-performed
Rüttimann's and we therefore elected to conduct a
controlled study to compare the two algorithms for
their ability to perform contrast corrections.
Histogram flattening can be thought of as a special
case of matching in which the desired histogram is flat
(i.e. equal numbers of pixels at each grey level). This
results in a CDF that follows the major diagonal.
Because flattening always matches to a CDF of fixed
shape, we hypothesized that flattening procedures with
both Rüttimann's and Castleman's algorithms would
result in superior contrast corrections than would the
corresponding matching algorithms that match to the
arbitrary shapes of the target CDF.
The aim of this study was to compare Rüttimann's
algorithm for matching (RM) and flattening (RF) and
Castleman's algorithm for matching (CM) and flatten-
ing (CF) for their ability to correct for contrast varia-
tions in dental digital images.
Materials and methods
Experimental model
Contrast variation in digital radiographic images may
occur because of changes in film exposure, film proces-
sing or radiograph digitization1"3'10. In our study, we
varied film exposure and held film processing and
radiograph digitization constant. We feel that this is
justified in that contrast variations attributable to film
processing and radiograph digitization are similar to
those attributable to film exposure.
A stepwedge was used to test for differences in
algorithm performance. The stepwedge consisted of
eleven 10 x 10 cm HVL aluminium attenuators (Nuc-
lear Associates, Carle Place, NY, USA) that changed
thicknesses in increments of either 0.1, 0.5 or 1.0 mm.
Each step was approximately 5-6 mm deep. Step-by-
step decreases in thickness varied from 37% to less than
2% (for example, the change from 2.7 mm to 1.7 mm
from step 5 to 4 represented a decrease in aluminium
thickness of 37%).
Image acquisition
The stepwedge was placed over Ektaspeed films (East-
man Kodak Co., Rochester, NY, USA) and exposed at
70 kVp, 7 mA and with a 30 cm FFD (focal film
distance), for 0.06, 0.12 and 0.25 s using a Heliodent
MD dental unit with a 12° anode angle and 2.1 mm of
aluminium filtration (Siemens Medical Systems,
Erlangen, Germany). To cover the full length of the
wedge, two films were needed. Six steps nearer the thin
end of the stepwedge were included on one film, and
six steps nearer the thick end of the stepwedge were
included on another film. Step 6 was included in both
exposures. In all, six films were exposed. The exposed
portions of the films were framed by lead blocks to
minimize scatter radiation, including backscatter.
43
Films were processed under highly controlled condi-
tions using Wing-Lynch one-shot processing with a
custom-made rotary tube (Wing-Lynch, Beaverton,
OR, USA). The developer was stored in a sealed
chemical tank filled with nitrogen to minimize oxida-
tion. Agitation was provided in a multidirectional
rotary pattern with a slight rocking action. The de-
veloper temperature was maintained at 20°C with a
water-jacketed processing trough. Films were proces-
sed using a universal rapid access developer, DC-4
(Christie, Toronto, Canada), and fixed with Rapid
Thiofix (Christie). The chemicals were discarded after
each batch of films was processed.
Radiographs were digitized at 50 μιη spatial resolu-
tion and 12-bit contrast resolution using a Molecular
Dynamics Personal Densitometer (Sunnyvale, CA,
USA). To avoid including the rounded corners of the
digital radiographic images, the images were cropped
to 654 x 256 pixels.
Grey-scale correction methods
The two methods of histogram matching and flattening
compared in this study both utilize cumulative grey-
scale distributions in which a histogram of the frequen-
cies of pixels at each grey-scale value is calculated, and
then the running sum of these is calculated to yield the
CDF 5 1 4 .
With the matching algorithm given by Rüttimann et
al.5, the cumulative sums of pixels at each grey scale in
the original image are matched to the corresponding
cumulative sums of the desired image, such that the
cumulative sum of the desired image is less than or
equal to the cumulative sum of the original image while
also being greater than the next lower bin of the
cumulative sum of the original image. Inherent in the
process is the combining of bins in the original image to
match the cumulative sums of the desired image.
For Castleman matching, the CDF of the input
(original) image is forced to match as closely as possible
the CDF of the target (desired) image (Figure 1). To
calculate the output grey level of an input grey level,
the value of input grey level is determined on the
original CDF. This value is used as input to the inverse
of the desired CDF, and this inverse grey-scale value of
the desired CDF is used as the output grey level (i.e. all
pixels in the image having the input grey level are
'mapped' to the output grey level). Castleman 14 pre-
sents a flat histogram with equal numbers of pixels at
each grey level except for the zero grey level, which has
no pixels. This results in a CDF that follows the major
diagonal (Figure 2). An alternative definition of a flat
histogram assumes equal numbers of pixels at each grey
level, including the zero grey level15. We conducted our
experiments with both types of histogram flattening but
found no overall difference in performance for these
methods. We present only results for histogram flatten-
ing performed using Castleman's concept of a flat
histogram.
The images created by exposures of 0.12 s were
considered to be the target images. The under- and
44 C.F. Hildebolt et al.

overexposed images created at 0.06 and 0.25 s, respec-

tively, were matched to the target images using Rütti-
mann's algorithm and Castleman's algorithm. These
Original CDF
same algorithms were also used to flatten the images.
α
3
> Our procedures resulted in four types of image:
or
<D Rüttimann-matched (RM), Castleman-matched (CM),
Desired CDF
Rüttimann-flattened (RF), and Castleman-flattened
(CF).
A program was written to sample 40 X 40 pixel areas
α of the resulting images. Samples were taken at the tops
ε and bottoms of each step of the stepwedge at the same
ο
Ζ locations for corresponding images (Figure 3). Mean
pixel grey-scale values were determined for each step.
For histogram matching, absolute differences in mean
input output
pixel grey-scale values between the unmodified target
0 1
Normalized Gray Level images (0.12 s exposure) and the histogram-matched
Figure 1 Castleman histogram matching. Both the original and
under- and overexposed images were determined. For
desired CDFs appear in the figure. To determine the output grey histogram-flattening, absolute differences in mean pix-
level of the input grey level, the value of the input on the original el grey-scale values between the flattened target images
CDF is determined (input arrow in the graph). This value is used as (0.12 s exposure) and the flattened under- and over-
input to the inverse of the desired CDF, and this inverse grey-scale
value (output arrow) of the desired CDF is used as the output exposed images were determined.
grey-scale value. Both the number of pixels and grey levels are
normalized to 1

Results
The optical densities for the steps in the optimally
exposed radiographs varied from 0.849 to 1.844. This
represented a range in grey-scale values of 995 of the
o>
σ possible 4096 grey-scale values in the 12-bit images.
Φ
Algorithm performance did not appear to be related to
Φ
>< stepwedge thickness nor exposure level, i.e. no pattern
α. could be detected for the absolute differences between
•ο the mean pixel grey-scale values for the steps of the
Φ
Μ
10 optimally exposed radiographic images and the corre-
ε sponding values for the histogram-matched images.
δ
2 Likewise, no pattern could be detected for the means
of the absolute differences between the histogram-
flattened target image and the histogram-flattened
input output under- and overexposed radiographic images. Figure 4
0 1 illustrates the lack of any pattern. It contains a plot of
Normalized Gray Level
differences for CM underexposed images against step-
Figure 2 Histogram flattening. With histogram flattening, the de- wedge thicknesses. Because of this lack of pattern,
sired CDF is flat. Input and output are calculated as described in summary statistics for algorithm performances were
Figure 1

TOP

BOTTOM

Figure 3 Sampling of a digital radiographic image of a stepwedge. Two radiographic films were used to cover the length of the stepwedge. The
image in the figure is of a Castleman-matched underexposed image of the thin end of the stepwedge. Six steps and the background (step 0) were
sampled for the images of the thin end of the stepwedge. Six steps were also sampled for the thick end of the stepwedge.
Comparison of contrast-correction methods 45

14η Table I Descriptive statistics for flattening and matching pro-

cedures. Algorithm performance was averaged for the stepwedge.
Values represent absolute differences in pixel grey-scale values
12 -
between target images and modified images

10- Correction method Minimum Maximum Median

>(0 8 - CF underexposed 2.6 221.9 69.8

Φ CF overexposed 0.4 193.9 70.5
(Ο
ο RF underexposed 2.6 221.9 69.8
ω
>. 6 Η RF overexposed 0.4 193.9 70.5
α RM underexposed 0.3 13.5 4.4
5
4Η RM overexposed 0.2 17.3 3.8
CM underexposed 0.2 13.5 4.4
2 - CM overexposed 0.4 17.0 4.2

0 1.0 1.1 1.6 1.7 2.7 3.7 4.7 4.8 5.3 5.4 6.4
Aluminum Thickness

Figure 4 Absolute differences between mean grey-scale values for

each step of the target image stepwedge and the CM underexposed S°
images of the stepwedge 1δ 1«
cu
g CO
®
£ £5
calculated for the images of the stepwedge. Because οα
the data values used for the summary statistics were not
normally distributed, means and standard deviations
were not good indicators of central tendency, and thus " 근 근 근 근 " "�
medians and percentiles were used (Table I and Figure 5). under-
exposed
over-
exposed
under-
exposed
over*
exposed
under-
exposed
over-
exposed
under-
exposed
over-
exposed
Box plots of the differences for the matching and
flattening procedures are presented in Figure 5. The Figure 5 Box plots of absolute grey-scale differences between target
images and histogram-matched and flattened images. The five hori-
four box plots to the left are for the CF and RF under- zontal lines of the box plots are for the 10th, 25th, 50th (median
and overexposed images. The four box plots to the value), 75th and 90th percentiles. Values above the 90th and below
right are for the RM and CM under- and overexposed the 10th percentiles are indicated by outlier dots
images. In all cases, the median differences between
the target images and the flattened images are greater 1 8

than 69 grey-scale values, which is approximately 7% of 1 6

the grey-scale range utilized for these images. For the
14
matched images, the median differences from target
images are less than five grey-scale intervals (less than £S ®3 1 2

1% of the grey-scale range). II 10

ε ®
Figure 6 is an enlargement of the RM and CM box S8 8
plots of Figure 5. Absolute differences in grey-scale Hs
b a 6
values between the target images and the RM under-
and overexposed images are shown on the left and 4
those for the CM corrections are shown on the right. 2
The algorithms showed essentially identical perform-
ance in correcting for contrast variations among im- RM RM CM CM
ages. The largest difference in grey-scale values for an underexposed overexposed underexposed overexposed

individual step between Rüttimann and Castleman Figure 6 Enlargement of Figure 5 for RM and CM images
matching was 0.4 of a grey-scale value.

grey-scale bins in the original CDF to match the

Discussion
To detect small changes in hard tissue densities in
longitudinal dental radiographic studies, some form of
contrast correction is essential. In 1986, a robust non-
parametric method of histogram matching was intro-
duced 5 . Now termed the Rüttimann method, it has
gained wide acceptance and has been found to produce
superior results to other contrast correction methods 10 .
Castleman produced a similar algorithm in 1979. Inhe-
rent in the use of Rüttimann's and Castleman's algo-
rithms for histogram matching is the combining of
cumulative sums of the desired CDF. This combining of
bins results in the loss of information in that it creates
additional zero-frequency bins. Both of these algo-
rithms can also be used for histogram flattening.
Because histogram flattening always matches to a CDF
of fixed shape as opposed to histogram matching which
matches CDFs of arbitrary shapes to one another,
we hypothesized that histogram flattening would out-
perform histogram matching for contrast correction.
In our study, the underexposed radiographs received
50% less radiation and the overexposed radiographs
received 100% more radiation than the target radio-
46
Gray Scale Values
Figure 7 Examples of histograms for the original underexposed
image, the desired target image, and the matched image after
histogram matching. The histogram for the matched image has
characteristics of both the original and target image histograms
graphs. As documented in Table I and in Figure 5, our
hypothesis that flattening procedures would result in
superior contrast corrections to those from matching
procedures alone is not supported. The average differ-
ence in the median values between flattened images
and target images was 66.0 grey-scale values, which
represents 6.6% of the grey-scale values used for the
target images. Individual differences per step were
large, as much as 222 grey-scale values, which repre-
sents 22% of the grey-scale range.
We speculate there are several reasons why flatten-
ing does not work well when compared with matching.
First, for flattening, two flattening procedures are
required (both the target histogram and the histogram
to be modified are flattened), whereas, for matching,
only one set of operations is needed. The extra proce-
dure with flattening introduces additional errors.
Second, the matched histogram takes on some of the
characteristics of the target histogram, yet it retains
many of the shape characteristics of the original histo-
gram (Figure 7). In performing a histogram flattening
procedure, the original and target histograms have very
different shapes from the flat histogram to which they
are being matched. Forcing the histograms to adapt to
flat shapes introduces additional errors. In summary,
histogram flattening is not appropriate for performing
contrast corrections in quantitative dental digital imag-
ing studies.
Rüttimann's and Castleman's algorithms showed
almost exactly the same performance in correcting
digital image contrast differences (Table I and Figure
6). The only difference that we could detect between
the two algorithms was in the use of CPU time. The
algorithms were run on a Sun Sparestation 10, Model
41 (Sun Microsystems Computer Corporation, Moun-
tain View, CA). Castleman matching used approx-
imately 3 s of CPU time whereas Rüttimann matching
used 25 s. However, we did not attempt to optimize the
performance of the Rüttimann algorithm and this may
account for the CPU time differences.
Histogram matching performed well. The average of
C.F. Hildebolt et al.
median differences in pixel grey-scale values between
the target images and the histogram-matched images
was 4.2, which represents about 0.4% of the grey-scale
range used (Table I). The largest difference was 17
grey-scale values, which represents 1.7% of the range
of grey-scale values. No pattern to the errors with
regard to corrections for underexposed or overexposed
images nor for thickness of aluminium could be de-
tected.
Rüttimann's and Castleman's algorithms performed
equally in correcting contrast on digital images, with
the largest difference in performance for a single step of
the stepwedge being only 0.4 of a grey-scale unit.
Although in our experiment Castleman's algorithm was
more than eight times faster than Rüttimann's, this
difference may be reduced by optimizing the Rütti-
mann algorithm. Moreover, the excellent results that
we obtained for histogram-matching contrast correc-
tions were obtained under highly controlled laboratory
conditions and are undoubtedly far better than could be
obtained in a clinical situation.
Acknowledgements
This study was supported in part by National Institute
of Dental Research (NIDR) projects DE08173 and
DE09861.
References
1. Ohki M, Okano T, Yamada N. A contrast-corrected method for
digital subtraction radiography. J Periodont Res 1988; 23: 277-
80.
2. Reddy MS, Jeffcoat MK. Digital subtraction radiography. Ad-
vances in dental imaging. Dent Clin North Am 1993; 37: 553-65.
3. Brägger U, Bürgin W, Marconi M, Häsler RU, Lang NP.
Influence of contrast enhancement and pseudocolor transforma-
tion on the diagnosis with digital subtraction images (DSI).
J Periodont Res 1994; 29: 95-105.
4. Fourmousis I, Brägger U, Bürgin W, Tonetti M, Lang NP.
Digital imaging processing: II. In vitro quantitative evaluation of
soft and hard peri-implant tissue changes. Clin Oral Implants Res
1994; 5: 105-14.
5. Rüttimann U, Webber R, Schmidt E. A robust digital method
for film contrast correction in subtraction radiography. J
Periodont Res 1986; 21: 486-95.
6. Brägger U, Pasquali L, Rylander H, Carnes D, Kornman KS.
Computer assisted densitometric image analysis in periodontal
radiography. A methodological study. J Clin Periodontal 1988;
15: 27-37.
7. Steffensen B, Pasquali L, Yuan C, Wood RC, Schoolfield JD,
Kornman KS. Correction of density changes caused by methodo-
logical errors in CADIA. J Periodont Res 1989; 24: 402-8.
8. Deas DE, Pasquali LA, Yuan CH, Kornman KS. The rela-
tionship between probing attachment loss and computerized
radiographic analysis in monitoring progression of periodontitis.
J Periodontal 1991; 62: 135-41.
9. Jeffcoat MK, Page R, Wannawisute A, Waite P, Palcanis K,
Cogen R. Use of digital radiography to demonstrate the poten-
tial of naproxen as an adjunct in the treatment of rapidly
progressing periodontitis. J Periodont Res 1991; 26: 414-21.
10. Fourmousis I, Brägger U, Bürgin W, Tonetti M, Lang NP.
Digital imaging processing: I. Evaluation of grey level correction
methods in vitro. Clin Oral Implants Res 1994; 5: 37-47.
11. Okano T, Ohki M, Mera T, Soejima H, Ishikawa I, Yamada N.
Comparison of contrast-correction methods
Quantitiative evaluation of proximal bone lesions using digital
subtraction radiography. Dentomaxillofac Radiol 1988; 17: 99-
103.
12. Robb RA, Barillot C. Interactive display and analysis of 3D
medical images. IEEE Trans Med Imaging 1989; 8: 217-26.
13. Robb RA, Hanson DP, Karwoski RA, Larson AG, Workman
EL, Stacy MC. ANALYZE: A comprehensive operator-
interactive software package for multidimensional medical image
display and analysis. Comput Med Imaging Graph 1989; 13:
47
433-54.
14. Castleman K. Digital Image Processing. New Jersey: Prentice
Hall, 1979: 90-3.
15. Lim JS. Two-dimensional Signal and Image Processing. New
Jersey: Prentice Hall, 1990: 453-9.
Address: Dr Charles F. Hildebolt, Mallinckrodt Institute of Radi-
ology, Washington University Medical Center, 510 South Kingshigh-
way Blvd, St. Louis, MO 63110, USA.