Computer Methods and Programs in Biomedicine 148 (2017) 1–11

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

journal homepage: www.elsevier.com/locate/cmpb

Automatic methods for alveolar bone loss degree measurement in

periodontitis periapical radiographs
P.L. Lin a,∗, P.Y. Huang b, P.W. Huang b
a
Department of Computer Science and Information Engineering, Providence University, 200, Taiwan Bld., Shalu, Taichung 43301, Taiwan
b
Department of Computer Science and Engineering, National Chung Hsing University, Taichung 40227, Taiwan

a r t i c l e i n f o a b s t r a c t

Article history: Background and objective: Periodontitis involves progressive loss of alveolar bone around the teeth. Hence,
Received 17 October 2016 automatic alveolar bone loss measurement in periapical radiographs can assist dentists in diagnosing such
Revised 14 May 2017
disease. In this paper, we propose an automatic length-based alveolar bone loss measurement system
Accepted 23 June 2017
with emphasis on a cementoenamel junction (CEJ) localization method: CEJ_LG.
Method: The bone loss measurement system first adopts the methods TSLS and ABLifBm, which we pre-
Keywords: sented previously, to extract teeth contours and bone loss areas from periodontitis radiograph images. It
Periodontitis radiographs then applies the proposed methods to locate the positions of CEJ, alveolar crest (ALC), and apex of tooth
Alveolar bone-loss measurement root (APEX), respectively. Finally the system computes the ratio of the distance between the positions of
CEJ localization
CEJ and ALC to the distance between the positions of CEJ and APEX as the degree of bone loss for that
Enamel dentin border detection
tooth. The method CEJ_LG first obtains the gradient of the tooth image then detects the border between
the lower enamel and dentin (EDB) from the gradient image. Finally, the method identifies a point on the
tooth contour that is horizontally closest to the EDB.
Results: Experimental results on 18 tooth images segmented from 12 periodontitis periapical radiographs,
including 8 views of upper-jaw teeth and 10 views of lower-jaw teeth, show that 53% of the localized CEJs
are within 3 pixels deviation (∼ 0.15 mm) from the positions marked by dentists and 90% have deviation
less than 9 pixels (∼ 0.44 mm). For degree of alveolar bone loss, more than half of the measurements
using our system have deviation less than 10% from the ground truth, and all measurements using our
system are within 25% deviation from the ground truth.
Conclusion: Our results suggest that the proposed automatic system can effectively estimate degree of
horizontal alveolar bone loss in periodontitis radiograph images. We believe that our proposed system,
if implemented in routine clinical practice, can serve as a valuable tool for early and accurate diagnosis
of alveolar bone loss in periodontal diseases and also for assessing the status of alveolar bone following
various types of non surgical and surgical and regenerative therapy. For overall system improvement, a
more objective comparison by using transgingival bone measurement with a periodontal probe as the
ground truth and enhancing the localization algorithms of these three critical points are the two major
tasks.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction degree of alveolar bone loss by examining radiographs has been a

Periodontitis involves progressive loss of the alveolar
bone around the teeth. Its diagnosis can be established from
(a) clinical examination by probing the gingiva (gums) using a
periodontal probe, or (b) radiographic examination by evaluating
the patient’s radiographs (X-ray films) to determine the amount
of alveolar bone loss around the teeth. Determining type and
∗
Corresponding author.
E-mail addresses: lan@pu.edu.tw (P.L. Lin), grances81@gmail.com (P.Y. Huang),
powhei.huang@msa.hinet.net (P.W. Huang).
difficult task for most dentists. Recently, two reports on validation
of a dental image analyzer tool to measure the degree of alveolar
bone loss are available in periodontitis research literature [1,2].
In [1], the degree of bone loss is measured horizontally by using
a Schei ruler [3] as shown in Fig. 1(a) to determine the ratio of
the distance between the position of the alveolar crest (ALC) (also
named as bone loss point BLP) and the position of the cemem-
toenamel junction (CEJ) to the distance between the CEJ and the
APEX, where ALC is the junction of the most apical extension of
the intrabony defect and the tooth, and the APEX is the apical of
the tooth root. Fig. 1(b) shows the measurement of length-based

http://dx.doi.org/10.1016/j.cmpb.2017.06.012
0169-2607/© 2017 Elsevier B.V. All rights reserved.
2 P.L. Lin et al. / Computer Methods and Programs in Biomedicine 148 (2017) 1–11
Fig. 1. Alveolar bone loss degree measurements based on length and defect angle: (a) Schei ruler (b) measured by length ratio: distal side (blue) 40 ∼< 50%, mesial side
(red) 30 ∼< 40%, (c) measured by defect angle. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
bone loss degree of a given tooth on both distal and mesial
sides, respectively. In [2], the degree of bone loss is measured
vertically and is determined by a defect angle between the two
lines that represent the root surface of the tooth involved and the
surface of the bone defect, shown as the lines formed by points
(i) (iii) and (ii) (iii) in Fig. 1(c), respectively. However, the critical
points/positions for either horizontal or vertical measurement are
manually marked by dentists. For automatic measurement of de-
gree of alveolar bone loss, bone loss areas as well as the positions
of CEJ, ALC, and APEX must firstly be localized automatically.
Very few studies on automatic bone-loss area localization and
teeth segmentation from periodontitis radiographs have been re-
ported in literature, although computer vision added feature or at-
tribute extraction from radiograph images for improving diagnosis
accuracy and aiding in clinical decisions are quite a few. For ex-
ample, in [4], intensity histogram, continuous wavelet transform,
and GLCM descriptors are combined for segmenting carotid artery
ultrasound images. In [5], grayscale features, local binary patterns,
and wavelet based features were combined for detecting the con-
dition coronary artery disease. In [6], textures of periapical radio-
graph images were used as a tool for dental implant treatment
planning.
We had presented an effective method for segmenting teeth
and extracting teeth contours from periapical radiographs in
[7] and a couple of methods for localizing alveolar bone loss areas
in periodontitis radiographs in [8–10]. Thus, positions of ALC and
APEX can be obtained from the located bone loss area and/or the
tooth contour. As for CEJ localization, no automatic methods have
been reported in literature up-to-date, to our best knowledge.
In this paper, we propose an automatic length-based alveolar
bone loss measurement system with emphasis on a CEJ localiza-
tion method. The CEJ localization method first obtains the gradient
from tooth images segmented from periodontitis radiographs then
detects the border between the lower enamel and dentin (EDB)
from the gradient image. Finally, the method identifies a point on
the tooth contour that is horizontally closest to the EDB. The bone
loss measurement system first adopts the methods TSLS [7] and
ABLifBm [10] to extract teeth contours and bone loss areas from
periodontitis radiograph images. It then applies the proposed CEJ,
ALC, APEX localization methods to obtain these three positions. Fi-
nally the system computes the ratio of the distance between the
positions of CEJ and ALC to the distance between the positions of
CEJ and APEX as the degree of bone loss for that tooth. For perfor-
mance assessment, 18 tooth images segmented from 12 periodon-
titis periapical radiographs, including 8 views of upper-jaw teeth
and 10 views of lower-jaw teeth will be tested. Experimental re-
sults will be presented both visually and quantitatively to demon-
strate effectiveness of the system.
The remainder of this paper is organized as follows. In
Section 2, the adopted teeth segmentation method TSLS and the
alveolar bone loss area localization method ABLifBm are briefly in-
troduced, followed by detail descriptions of the proposed cemen-
toenamel localization method and the length-based alveolar bone
loss measurement system. In Section 3, experimental results and
performance assessment of the CEJ localization method and the
bone loss measurement system are presented and discussed, re-
spectively. Finally, conclusions are given in Section 4.
2. Materials and methods
In this section, we first briefly describe the methods and sys-
tems adopted in the proposed system. We then describe our pro-
posed system and methods in detail.
2.1. Study samples
Twelve periodontitis periapical radiograph images of various
alveolar bone loss degrees are used in this study. All images are
provided by Chung Sun Medical University Hospital in Taichung,
Taiwan.
2.2. Adopted methods and systems
2.2.1. Teeth segmentation using local singularity: TSLS
TSLS [7] is a threshold segmentation method that can ef-
fectively segment each tooth from a periapical radiograph. The
method first applies an adaptive power law transformation to the
image to reduce uneven illumination problem. It then recognizes
each tooth in the radiograph by thresholding the bilateral filtered
point-wise singularity image, where the singularity of each point
is characterized by the Hölder exponent α at the point defined as
ln(μi (x, y ))
αi (x, y ) = , i = 1, 2, 3... (1)
ln(i )
where μi (x, y) is the amount of capacity measure within the ob-
served box with size = i centered at the pixel (x, y). Some of the
commonly used capacity measures are: max, min, iso, and sum of
the gray-level intensity at points within the measured box. Finally,
 P.L. Lin et al. / Computer Methods and Programs in Biomedicine 148 (2017) 1–11
Fig. 2. Segmentation results of 4 periapical radiograph images: (x − 1) original images, (x − 2) segmented teeth and contours.
the contour of each tooth is extracted by tooth isolation and shape
refinement. Fig. 2 shows the segmentation result of four periapical
radiograph images, including two lower-jaw tooth views and two
upper-jaw tooth views.
2.2.2. Alveolar bone loss area localization method- ABLIFBM
The method ABLifBm [10] is a threshold segmentation method
for alveolar bone loss area localization. It uses a hybrid feature
fused of both intensity and texture measured by the H-value of
fractional Brownian motion (fBm) model, where the H-value is the
Hurst coefficient in the expectation function of a fBm curve (inten-
sity change) and is directly related to the value of fractal dimen-
sion. That is, the H-value can be obtained from
log(E (Ir )) = H log(r ) + K (constant ) .
where
Ir = |I (x2, y2 ) − I (x1, y1 )|
 edges.
r = (x2 − x1 )2 + (y2 − y1 )2
and fractal dimension FD = 3−H.
Adopting leave-one-out cross validation training and test-
ing mechanism, ABLifBm trains weights for both features using
Bayesian classifier and transforms the radiograph image into a fea-
ture image obtained from a weighted average of both features. Fi-
nally, by Otsu’s thresholding, it segments the feature image into
normal and bone loss regions.
Fig. 3 shows the alveolar bone loss areas (marked in red shade)
detected by ABLifBm in four periodontitis radiographs.
2.2.3. ISO 3950 notation tooth numbering system
ISO 3950 notation tooth numbering system [11], as shown in
Fig. 4, is devised by the world health organization notation system
and is widely used by dental professionals internationally. The sys-
tem uses a two-digit numbering system in which the first number
represents a tooth’s quadrant (e.g., upper right, upper left, lower
left, lower right) and the second number represents the number
of the tooth from the midline of the face. The orientation of the
illustration is according to dentist’s view, i.e., the notation “left”
corresponds to patients’ right-side face, and the notation “right”
corresponds to patients’ left-side face, respectively. The permanent
3
teeth are divided into four quadrants, where teeth in the upper
right quadrant begin with the number “1”, teeth in the upper left
quadrant begin with the number “2”, teeth in the lower left quad-
rant begin with the number “3”, and teeth in the lower right quad-
rant begin with the number “4”.
2.2.4. Bilateral filtering
Bilateral filter, which consists of domain filter and range filter,
is one of the most popular methods to smooth images while pre-
serving edges. The idea is as follows [12].
Two pixels in an image can be close to each other, or they can
be similar to each other. Closeness refers to vicinity in the spatial
domain, while similarity referring to vicinity in the image value
range. Domain filtering enforces closeness by weighing pixel values
(2) with coefficients that fall off with distance, whereas range filtering
averages pixel values with weights that decay with dissimilarity.
By combining both domain filter and range filter together, one can
(3) obtain a bilateral filter that can smooth images while preserving
The mathematical formulas for bilateral filter (BF) are as follows
(4)
1  
BFσd ,σr (I ) = P  ∈ Gd (P − P )Gt (IP − IP  )IP (5)
k (P )

K (P ) = P  ∈ Gd (P − P )Gr (IP − IP ) (6)
x2
Gd (x ) = exp(− ) (7)
σd2
x2
Gr (x ) = exp(− ) (8)
σr2
where Gd and Gr are Gaussian functions under the control of vari-
ance σ d for domain filter and σ r for range filter, respectively, P is
the pixel location, P is the neighborhood pixel location, IP is the
intensity value of the pixel at location P, IP  is the intensity value
of the pixel at location P ,  is the neighborhood, and K(P)−1 is a
normalized similarity function.
2.3. The proposed cementoenamel junction localization method
The cementoenamel junction (CEJ) is an anatomical border
identified on a tooth. It is the location where the enamel and the
4 P.L. Lin et al. / Computer Methods and Programs in Biomedicine 148 (2017) 1–11
Fig. 3. Four periapical perdontitis radiographs with the localized bone loss areas (red shade areas) and ground truth (areas within black contours): (x − 1) original images,
(b) (x − 2) localized and ground truth bone loss areas. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this
article.)
Fig. 4. ISO 3950 notation tooth numbering system [11].
cementum meets. Enamel is the part that covers the anatomical
crown of a tooth. It can be divided to upper enamel and lower
enamels on distal and mesial sides, where the distal side is the
side towards the gingiva beyond the tooth furthest from the ante-
rior midline (i.e., away from the midline in a dental arc), and the
mesial side is the side towards the anterior midline in a dental
arch. The cementum, which covers the anatomical root of a tooth,
is invisible in dental radiographs. However, it is right at outside of
the root contour, thus, the location of CEJ can be identified as the
junction of the lower enamel and tooth-root contour.
Our proposed cementoenamel junction method CEJ_LG can
be proceeded in three major steps: (A) image enhancement, (B)
enamel and dentin border (EDB) detection, and (C) EDB-tooth con-
tour junction localization.
2.3.1. Image enhancement
Dental radiographs are often subject to noise, low contrast, and
uneven illumination. As shown in Fig. 5(a), intensity differences
between the lower enamel and dentin at the left side (mesial side
of this tooth) are quite insignificant, and intensity varies gradually
from the mesial side to the distal side. In other words, both the
EDB and the CEJ are difficult to be identified at the mesial side of
this tooth without enhancing the image. The proposed image en-
hancement procedure is as follows.
1. Denoising:
Apply bilateral filtering [12] to reduce noise while preserving
edge information.
2. Contrast stretching:
Apply power law transformation [13] to stretch the contrast be-
tween enamel and dentin pixels.
Fig. 5(a) shows the upper half of a tooth image and Fig. 5(b)
shows the result of (a) after the image enhancement. Notice that
changes of the intensity between enamel and dentin in (b) are
much significant than those in (a) and the border between lower
enamel and dentin at the mesial side becomes wider and darker,
which makes allocation of this border easier.
2.3.2. Enamel and dentin border (EDB) detection
After image enhancement, the intensity differences between
enamel and dentin pixels become more significant, i.e., the gra-
dient values of the EDB pixels are greater than those of enamel
pixels and dentin pixels. Thus, we can find the EDB as the pixels
in the border neighborhood having the largest gradient value. The
EDB tracking procedure is as follows.
1. Obtain the gradient image Gx using Sobel horizontal filter [13].
2. Identify the minimum bounding rectangle (MBR) of the tooth
on the gradient image then discard the lower half and retain
the upper half.
 P.L. Lin et al. / Computer Methods and Programs in Biomedicine 148 (2017) 1–11
Fig. 5. An example of CEJ-LG on a tooth image: (a) the original image, (b) the image after denoising and contrast stretching, (c) the horizontal gradient image, (d) modified
gradient image and the row rs on the mesial and the distal sides, (e) tracked EDB (marked in black) on the mesial and distal sides, (f) located CEJ (red) and ground truth
(green) on mesial and distal sides. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
3. Remove both the top left and top right tooth boundaries from
the upper-half gradient image by modifying the Gx .
a. Select the left half as ROI and set the positive gradient to 0
and the negative gradient to positive.
b. Select the right half as ROI and set the negative gradient to
0.
4. Find the starting pixel position of EDB (rs , cs ) within each ROI.
a. Find rs as n rows below the row r that has the maxi-
mum horizontal intensity integral projection value of the
enhanced image within ROI.
b. Find cs as the column where the gradient at point (rs , cs ) is
the first positive value scanning from the outside of tooth
boundary to tooth.
Note that we set the starting row rs as n rows down from the
row r so that all pixels in row rs are completely within the
lower enamel.
5. Locate the entire border pixels of EDB within each ROI.
a. Start at the pixel (rs , cs ) and find the next EDB pixel in the
row below as the one having the maximum gradient among
the three pixels centered at (rs + 1, cs )
b. Continue finding the next EDB pixels until the tooth bound-
ary or the bottom of the ROI is reached.
Figs. 5(c)–(e) show the results of steps 1, 4, 5 of EDB detec-
tion on the tooth image in (b), where (c) is the horizontal gradi-
ent image, (d) is the modified gradient images of the ROIs and the
starting row rs for tracking the EDB on the mesial (left) and distal
(right) sides, respectively, (e) is the tracked EDB (marked in black)
on the mesial and distal sides. Notice that both the left and right
tooth boundaries have disappeared in (d) so that the first signif-
icant curve scanning from outside the tooth will be EDB. Notice
also that the tracked EDB on the mesial side is quiet similar to the
bright curve appears at the left in (d). However, a big part of the
tracked EDB at the right side is not shown in (d) because the gra-
dient magnitude of the missing part is too small to be visible.
2.3.3. EDB and tooth-contour junction localization
Now that we had obtained the EDB so we should be able to lo-
cate the CEJ as the intersection of tooth boundary and EDB. How-
ever, the detected EDB may not intersect with the tooth contour
5
due to imperfect image quality, as illustrated in Fig. 5(e). Thus, in-
stead of identifying CEJ as a point on EDB intersecting with tooth
contour, we identify it as a point on the tooth contour that is hor-
izontally closest to EDB. The localization procedure is as follows.
1. Calculate the horizontal distance between the tracked EDB and
the tooth contour row-by-row.
2. Identify the row k that has the minimum distance.
3. Identify the CEJ location as the position of the point on the
tooth contour at the row k.
Fig. 5(f) shows the localized CEJs of the tooth (the center of the
red circles) in Fig. 5(a) on the mesial and distal sides. Notice that
the localized CEJ on the distal side is quite accurate although the
tracked EDB is not.
Although the procedure we describe above is for lower-jaw
teeth, it can be used also for upper-jaw teeth by simply turning the
images upside down. Thus, the proposed CEJ localization method
can be summarized as the following algorithm.
2.3.4. Algorithm of CEJ localization method
Algorithm: CEJ_LG
Input: an isolated lower periapical tooth image I
Output: CEJ locations on distal and mesial sides
1. Apply image enhancement to I and obtain the enhanced image
I .
a. Denoise the image using (5)–(8) with σ d = 1 and σ r = 0.1,
respectively, and with mask size 7 × 7.
b. Stretch the contrast between enamel and dentin pixels using
x−a γ
y=( ) (9)
b−a
with γ = 2.5, [a, b] being the intensity range of I.
2. Calculate the horizontal gradient image Gx using
 
−1 0 1
Gx = −2 0 2 ∗ I (10)
−1 0 1
where ∗ denotes 2-dimensional convolution operation.
3. Identify MBR of the tooth on the image I and retain the upper
half MBR.
6 P.L. Lin et al. / Computer Methods and Programs in Biomedicine 148 (2017) 1–11

Fig. 6. An example of the APEX and ALC localization: (a) a binarized segmented tooth, (b) the tooth contour and located apex (center of the red circle) (c) the binarized
Alveolar bone loss area along the tooth, (d) the located ALC (center of the red circle). (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

Fig. 7. Twelve periodontitis periapical radiographs for CEJ detection and alveolar bone loss degree measurement experiments.

4. Select the left-half of the MBR as ROI and modify the gradient d. Find cs where G(rs , cs ) > 0 and G(rs , k) = 0 , 0 <k <cs for left
image in the ROI by letting ROI and cs < k < w for right ROI, w is the width of ROI.
 6. Locate all border pixels of EDB and save in B for each ROI.
−Gx if Gx < 0
G= (11) a. Let i = 0, r = rs , c = cs
0 otherwise
b. Bi = (r, c) and i = i + 1
and select the right-half of the MBR as ROI and let c. Set r = r + 1 and find c where G(r, c ) = arg max G(r, k )
 c−1≤k≤c+1
Gx if Gx > 0 d. Repeat b and c until i > h
G= (12)
0 otherwise 7. Identify the location of CEJ
5. Find the starting pixel position of EDB (rs , cs ) for each ROI. a. Compute
a. Calculate the horizontal intensity integral projection of I di = Ci − Bi ,i = 1, ..., h (15)
within ROI, P(k), 1ࣚkࣚh, where h is the height of ROI.
b. Find the row r where where Bi is the ith point of the EDB and Ci is the tooth con-
tour position at the same row.
P (r ) = arg max P (k ). (13) b. Find k where dk = arg min di .
1 ≤k≤ h
1≤i≤h
c. Set c. Identify CEJ location as C rs + k -1 , a point on the tooth contour
at row rs +k-1.
rs = r + n(0 < n < 20) (14)
 P.L. Lin et al. / Computer Methods and Programs in Biomedicine 148 (2017) 1–11 7

2.4. Length-based alveolar bone loss degree measurement Table 1

Location deviation between detected DEJ and ground truth.

Based on the measuring method presented in [1], the length- IM# Tooth index Location deviation in pixels
based alveolar bone loss degree (ABLD) can be measured using two Distal site Mesial site
metrics: BL and TR, where BL is the distance between the positions
C1 23 – 3.6
of cementoenamel junction (CEJ) and alveolar bone crest (ALC), C5 15 7.1 5.1
and TR is the distance between the positions of CEJ and apex of C5 16 1.1 1.0
the tooth (APEX). The formulas are as follows. C6 15 0.0 6.3
C6 16 4.5 11.2
ABLD = (BL/T R ) × 100% (16) C6 17 2.0 5.4
C8 24 – 5.0
C12 45 2.8 2.2
BL = CEJ − ALC  (17) C13 36 1.4 2.2
C13 37 2.0 8.5
C14 46 8.6 2.2
T R = CE J − AP E X  (18) C17 35 – 3.0
C17 36 – 2.2
C19 35 – 1.4
2.4.1. APEX and ALC localization C20 27 1.0 5.0
Since APEX is at the apical of tooth root and ALC is at the apical C22 45 1.0 4.5
of alveolar bone loss areas along tooth root, we can find the APEX C23 44 6.3 1.4
C23 45 – 5.4
as the lowest (highest) pixel position on the contour of each lower- Summary 30 CEJs Mean ± std. deviation = 3.8 ± 2.67
jaw (upper-jaw) tooth and find ALC as the lowest pixel position
of the intersection of bone loss area and tooth contour. Thus, our
proposed APEX and ALC localization procedure is as follows. named based on ISO 3950 notation numbering system as shown in
1. Obtain the APEX of each tooth in the radiograph Fig. 4 [11]. For example, the tooth “16” stands for the sixth tooth
from the midline of the face in the upper right jaw, and the tooth
a. Apply the teeth segmentation method TSLS [7] to obtain
“23” stands for the third tooth away from the midline of the face
each segmented tooth in the radiograph.
in the upper left jaw, the tooth “35” stands for the fifth tooth from
b. Binarize each tooth image and apply Moore boundary track-
the midline of the face in the lower left jaw, and the tooth “45”
ing algorithm [14] to get the tooth contour.
stands for the fifth tooth from the midline of the face in the lower
c. Locate the lowest (highest) position of each tooth contour
right jaw, etc.
and identify it as the apex of lower-jaw (upper-jaw) tooth.
2. Obtain the ALC of each tooth 3.1. Performance assessment
a. Apply the bone loss area localization method ABLifBm
[10] to obtain bone loss areas of the associated tooth in the Performance of CEJ localization is assessed both visually and
radiograph. quantitatively. For visual assessment, both the located CEJ posi-
b. Binarize the bone loss area of the associated tooth and apply tions and the ground truth CEJ positions are marked on the same
morphological dilation so that the bone loss area overlaps tooth image. For quantitative assessment, deviation between the
with the tooth contour. detected and the ground truth CEJ positions is calculated in terms
c. Find the overlapping pixels of the dilated bone loss area and of number of pixels.
the tooth contour. Performance of alveolar bone loss degree measurement is as-
d. Identify the ALC as the bottom (top) position of the overlap- sessed quantitatively; however, the detected and the ground truth
ping pixels for lower- (upper-) jaw tooth. positions of ALC and APEX associated with each tooth are also
marked on the same image for visual references. For quantitative
Fig. 6 shows an example of the APEX and ALC localization pro- assessment, deviation between the bone loss degrees measured
cedure, where (a) is the binary image of a segmented tooth, (b) is based on the positions of localized CEJ, ALC and APEX and the po-
the contour of this tooth and the located APEX (center of the red sitions marked by dentists is used as the metric.
circle), (c) is the binary image of the localized alveolar bone loss
area along the tooth, and (d) is the located ALC (marked in red 3.2. Results
circle).
3.2.1. CEJ localization
3. Experiments and results The experimental results of 8 upper-jaw teeth and the results of
10 lower-jaw teeth are shown in Figs. 8 and 9, respectively, where
Two experiments are conducted to test effectiveness of our the center of each red circle is the located CEJ position and the
proposed CEJ localization method and the alveolar bone loss de- center of each green circle is the ground truth position marked by
gree measurement, respectively. In the first experiment, we test dentists. The index “Cx-yy” below each image refers to the image
18 tooth images segmented from 12 periodontitis periapical radio- “Cx” and tooth number “yy” in Fig. 7, respectively. Table 1 lists the
graphs with size of approximate 868x650 pixels each (8 in the location deviation in pixels between the located CEJ position and
upper jaw and 10 in the lower jaw), where the CEJs of these the ground truth CEJ position of all 18 teeth on distal and mesial
18 teeth in the radiographs are detectable. In the second experi- sides, respectively.
ment, we use the same set of tooth images to evaluate the per- From Table 1, we can find that out of 30 located CEJs, 16 po-
formance of the proposed alveolar bone loss degree measurement. sitions have deviation  <= 3 pixels, 13 positions have 3 <  <= 9
Fig. 7 shows the 12 periodontitis radiograph images used for ex- pixels, and 1 position has 9 <  <= 12 pixels. In other words, 53%
periments, where the index “Cx” below each image refers to the of the localized CEJs are within 3 pixels deviation from the po-
image number “x” in our database. Each tooth number “y1 y2 ” in sitions of the ground truth, and 90% have deviation less than 9
the image indicates that the tooth is in the “y1 ” quadrant and is pixels. As shown in Fig. 7 image C5, the length of the tooth num-
the “y2 ” numbered tooth from the midline of the face, which is bered 15 is approximate 470 pixels. Given a normal premolar tooth
8 P.L. Lin et al. / Computer Methods and Programs in Biomedicine 148 (2017) 1–11

Fig. 8. Results of CEJ detection in 8 upper-jaw tooth images: the centers of each red and green circle are the detected and ground truth CEJ locations, respectively. Note: for
teeth # 1x, the left side is the distal site and the right side is the mesial site, and for teeth # 2x, the left side is the mesial site and the right side is the distal site.

size of approximate 21–23 mm, the resolution of the tested image Table 2
Alveolar bone loss degree measured by the proposed system vs. ground
is about 0.045 ∼ 0.049 mm/pixel. Thus we can state that 53% of the
truth.
detected CEJ locations have deviation less than ∼ 0.15 mm, and 90%
have deviation less than ∼ 0.44 mm. IM# Tooth index ABLD (in %)

Distal site Mesial site

3.2.2. Alveolar bone loss degree measurement ours GT  ours GT 

Fig. 10 shows the positions of APEX, ALC and CEJ of the dis- C1 23 – – – 70.4 44.7 25.7
eased teeth in 11 periodontitis radiographs by shapes of diamond, C5 15 69.3 52.0 17.3 – – –
C6 15 69.7 86.9 17.2 81.0 73.2 7.8
square and circle, respectively, where red shapes represent posi-
C6 16 78.7 56.6 22.1 93.2 93.2 21.3
tions located by our proposed method and green shapes represent C6 17 46.7 53.4 6.7 49.7 41.4 8.3
the positions of the ground truth. Note that C23-45 used for CEJ C8 24 – – – 35.9 23.2 12.7
localization is missing in the figure, as no alveolar bone loss area C12 45 53.1 44.7 8.4 48.1 47.8 0.3
associated with the tooth 45 in C23 was identified by dentists. C13 36 – – – 52.7 49.0 3.7
C13 37 – – – 48.2 46.2 2.0
Table 2 lists degrees of alveolar bone loss measured by our pro-
C14 46 92.7 81.4 11.3 80.9 76.2 4.7
posed system and ground truth that is calculated based on the po- C17 35 – – – 48.7 23.3 25.4
sitions drawn by dentists, as well as bone loss degree deviation C17 36 – – – 44.8 23.4 21.4
() between the two measurements. Notice that out of 21 bone C19 35 – – – 69.6 66.0 3.6
C20 27 61.3 35.5 25.8 42.5 40.1 2.4
loss degree measurements, the average deviation is 9.5%, where 11
C22 45 24.6 25.1 0.5 – – –
measurements have deviation ࣚ 10%, 4 measurements have 10% <
deviation ࣚ 20%, and 6 measurements have 20% < deviation ࣚ 25%, Summary: mean  = 9.5%, #(ࣚ 10%) = 11, #(10% <  ࣚ 20%) = 4,
#(20% <  ࣚ 25%) = 6
respectively. In other words, we can see that more than half of the
alveolar bone loss degree measurements using our system have
deviation less than 10% from the ground truth, and all measure- truth, where the ground truth are calculated based on the locations
ments using our system are within 25% deviation from the ground marked by dentists.
 P.L. Lin et al. / Computer Methods and Programs in Biomedicine 148 (2017) 1–11 9

Fig. 9. Results of CEJ detection in 10 lower-jaw tooth images: the centers of each red and green circle are the detected and ground truth CEJ locations, respectively. Note:
for teeth # 3x, the left side is the mesial site and the right side is the distal site, and for teeth # 4x, the left side is the distal site and the right side is the mesial site. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

4. Analysis and discussion
Observing the images of the three worst results in the test of
CEJ localization, i.e., C6-16 mesial side, C14-46 distal side, and C13-
37 mesial side, we can find that (a) a big portion of the enamel in
C6-16 mesial side is missing and the portion visible in the image is
too thin so that the enamel dentin border is difficult to be tracked
accurately; (b) the enamel in C14-46 distal side has in fact decayed
so that the detected CEJ is at higher position than the true CEJ; (c)
the shadow around the mesial side CEJ of C13-37 causes the lower
part of the tracked EDB drifting away from the tooth boundary and
results in the detected CEJ at higher location.
Observing alveolar bone loss degree measurement results in
Table 2 and the CEJ, APEX, and ALC localization results shown in
Fig. 10, we found that four out of the six measurements with devi-
ation > 20% (i.e., C1-23, C17-35, C17-36, and C20-27) are all due to
unsatisfactory localization of the ALC/APEX. Only C6-16 is due to
unsatisfactorily located CEJ on distal side and APEX on mesial side,
respectively.
Thus, more accurate locations of ALC and APEX besides CEJ are
must for improving the system performance. However, positions
of these points are sometimes difficult to be accurately localized,
even by dentists, because of unsatisfactory radiographs’ quality. In
[15], the authors proposed a similar method to automatically eval-
uate the quantitative alveolar bone resorption (ABR). They applied
the method to 30 patient cases with alveolar bone resorption of
different stages and indicated that the performance still needs fur-
ther improvement. On the other hand, “the ground truth (gold
standard) for bone level measurement can also be transgingival
10 P.L. Lin et al. / Computer Methods and Programs in Biomedicine 148 (2017) 1–11

Fig. 10. Detected CEJ (red circle), ALC (red square), and APEX (red diamond) vs. ground truth CEJ (green circle), ALC (green square), and APEX (green diamond) of all teeth
used for bone loss degree measurement. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

bone measurement with a periodontal probe1 ”, which is “repeat-
able to within 1 mm better than 90% of the time” [16]. For over-
all system improvement, a more objective comparison by using
transgingival bone measurement with a periodontal probe as the
ground truth, instead of using the locations marked by dentists,
1
As suggested by the anonymous reviewer.
and enhancing the localization algorithms of these three critical
points are the two major tasks.
5. Conclusions
We presented an automatic length-based alveolar bone loss
measurement system with emphasis on a CEJ localization method.
We tested the system with 18 tooth images segmented from 12
 P.L. Lin et al. / Computer Methods and Programs in Biomedicine 148 (2017) 1–11
periodontitis periapical radiographs, including 8 upper-jaw teeth
and 10 lower-jaw teeth. Experimental results in terms of the posi-
tions of localized and ground truth CEJ, ALC, and APEX are marked
in the same image for visual evaluation. The quantitative results
show that 53% of the localized CEJs are within 3 pixels deviation
(∼ 0.15 mm) from the positions marked by dentists, and 90% have
deviation less than 9 pixels (∼ 0.44 mm). For degree of alveolar
bone loss measurement, more than half of the measurements us-
ing our system have deviation less than 10% from the ground truth,
and all measurements using our system are within 25% deviation
from the ground truth.
In our future study, we will improve both the algorithms of
teeth segmentation and alveolar bone loss area localization for
more accurate locations of APEX and ALC, respectively. Meanwhile,
we will compare the measurement with the transgingival bone
measurement done by dentists with a periodontal probe.
We believe that our proposed system, if implemented in routine
clinical practice, can serve as a valuable tool for early and accurate
diagnosis of alveolar bone loss in periodontal diseases and also for
assessing the status of alveolar bone following various types of non
surgical and surgical and regenerative therapy.
Acknowledgment
The authors would like to thank to Chung-Shan Medical Uni-
versity Hospital in Taichung, Taiwan for providing the radiograph
images and to Dr. Hsu for marking the ground truth. The authors
would also like to thank to the anonymous reviewers for provid-
ing with the suggestion of bone loss measurement gold standard.
The research was partially supported by Ministry of Science and
Technology, ROC under grant #: MOST 103-2221-E-126 -006 -MY2.
References
[1] W.J. Teeuw, L. Coelho, A. Silva, C.J. van der Palen, F.G. Lessmann, U. van der
Velden, B.G. Loos, Validation of a dental image analyzer tool to measure alve-
olar bone loss in periodontitis patients, J. Periodontal Res. 44.1 (2009) 94–102.
11
[2] R.P. Moutinho, L. Coelho, A. Silva, A.J. Lobo Pereira, M. Pinto, I.P. Baptista, Vali-
dation of a dental image-analyzer tool to measure the radiographic defect an-
gle of the intrabony defect in periodontitis patients, J. Periodontal Res. 47 (6)
(2012) 695–700.
[3] D.A. Miles, M.R. Razzano, The future of digital imaging in dentistry, Dent. Clin.
North Am. 44 (2) (20 0 0) 427–438.
[4] M. Hassan, A. Chaudhry, A. Khan, J.Y. Kim, Carotid artery image segmentation
using modified spatial fuzzy c-means and ensemble clustering, Comput. Meth.
Programs Biomed. 108 (3) (2012) 1261–1276.
[5] U.R. Acharya, S.V. Sree, M. Muthu Rama Krishnan, N. Krishnananda, S. Ranjan,
P. Umesh, J.S. Suri, Automated classification of patients with coronary artery
disease using grayscale features from left ventricle echocardiographic images,
Comput. Meth. Programs Biomed. 112 (3) (2013) 624–632.
[6] D.R. Dias, R.M. Costa, C.R. Leles, P.M. Azevedo-Marques, R.F. Ribeiro-Rotta,
M.B.V. Mundim, Intraoral radiographs texture analysis for dental implant plan-
ning, Comput. Meth. Programs Biomed. 136 (2016) 89–96.
[7] P.L. Lin, P.Y. Huang, P.W. Huang, H.C. Hsu, C.C. Chen, Teeth segmentation
of dental periapical radiographs based on local singularity analysis, Comput.
Meth. Programs Biomed. 113 (2014) 433–445.
[8] P.-L. Lin, P.-Y. Huang, P.-W. Huang, H.C. Hsu, P. Chen, Alveolar bone loss area
localization in periapical radiographs by texture analysis based on fBm model
and GLC matrix, in: Proc. IEEE Int. Sym. on Bioelectronics and Bioinformatics,
Taiwan, 2014 April 11-14.
[9] P.-L. Lin, P.-Y. Huang, P.-W. Huang, An automatic lesion detection method for
dental X-ray images by segmentation using variational level set, in: Proc. IEEE
Int. Conf. on Machine Learning and Cybernetics (ICMLC), 2012.
[10] P.L. Lin, P.W. Huang, P.Y. Huang, H.C. Hsu, Alveolar bone loss area localization in
periodontitis radiographs based on threshold segmentation with a hybrid fea-
ture fused of intensity and the H-value of fractional Brownian motion model,
Comput. Meth. Programs Biomed. 121 (2015) 117–126.
[11] ISO 3950:2009 Dentistry —Designation system for teeth and areas of the oral
cavity, https://www.iso.org/standard/41835.html.
[12] C. Tomasi, R. Manduchi, Bilateral filtering for gray and color images, Computer
Vision, Sixth International Conference on, IEEE, 1998.
[13] R.C. Gonzales, R.E. Woods, Digital Image Processing, second ed., 2002.
[14] G.A. Moore, Automatic scanning and computer processes for the quantitative
analysis of micrographs and equivalent subjects, Pictorial Pattern Recognit.
(1968) 275–326.
[15] M. Zhang, A. Katsumata, C. Muramatsu, T. Hara, H. Suzuki, H. Fujita, An auto-
matic early stage alveolar–bone-resorption evaluation method on digital dental
panoramic radiographs, in: SPIE Proceedings, 9035, Medical Imaging, 2014.
[16] M.K. Jeffcoat, M.S. Reddy, Advances in measurements of periodontal bone and
attachment loss, in: R.V. Faller (Ed.), Assessment of Oral Health, Monographs
in Oral Science, 17, Karger, Basel, 20 0 0, pp. 56–72.