Computer Aided Diagnosis for Osteoporosis Based on Trabecular Bone Analysis Using Panoramic Radiographs

abstract Background: BLOCKINpanoramic BLOCKINradiographs.


introduction
The number of hip fractures because of osteoporosis was rising from about 1.3 million in 1990 worldwide. Moreover, it was estimated to be 4.5 million in 2050. 1 The U.S. Surgeon General reported that if there was no serious handling until 2020, half of Americans were predicted to have osteoporosis. 2 One parameter should be measured to determine whether someone has osteoporosis or not is called the bone mineral density (BMD), measured on lumbar spine and femoral neck.
The commonly used scanner for measuring bone is dual-energy x-ray absorptiometry (DXA). However, DXA is very expensive and not every hospital has DXA even in developed countries. 3 On the other hand, postmenopausal women rarely visit medical expert to diagnose osteoporosis. Postmenopausal women would realize that they had osteoporosis after any bone fractures caused by an accident.
Postmenopausal women may have greater opportunity to visit dentists for treatment of dental caries and periodontal disease than to visit medical professionals for diagnosis of osteoporosis. A large number of panoramic radiographs were taken for diagnosis of teeth and jaws in general dental practice.
Trabecular bone is one of important parts of panoramic radiographs. The trabecular bone pattern may be analyzed visually by experts or with computer-aided methods to estimate the probability of having osteoporosis and predict the risk of future fractures. 4 Degrees of inter-examiner and intraexaminer agreement of visual assessment of the trabecular pattern are also expected to be relatively low because the trabecular pattern of the jaws is more diverse than that of the general skeleton, such as the vertebrae and proximal femur. 5 However, other researchers analyzed trabecular bone pattern using customized image analysis software. [6][7][8][9] This research proposed a method to analyze trabecular bone tissue using multiscale line operator on Gaussian pyramid. 10 The line strength of trabeculae that has similar orientation with the root of teeth was measured on both left and right sides of the mandible. High correlation between both experimental results and BMD assessed by DXA scanner proves the effectiveness of this method. 10 The purposes of this study were to develop a computeraided system for measuring trabecular bone line strength on panoramic radiographs in identifying postmenopausal women with osteoporosis and to clarify the diagnostic efficacy of the system. The density was measured with respect to line strength of segment structure which has similar orientation with the root of tooth. This research also determined the threshold value of line strength considered as the osteoporosis sign.

materials and methods
There were 531 women for DXA measurement between 1996 and 2001, 100 postmenopausal women aged 50 years or older with no previous history on osteoporosis (mean 59.6 years; range 50-84 years) were randomly recruited for this study. None of the subjects had metabolic bone disease (hyperparathyroidism, hypoparathyroidism, Paget's disease, osteomalacia, renal osteodystrophy, or osteogenesis imperfecta), cancers with bone metastasis, or significant renal impairment or were taking medication that affect bone metabolism, such as estrogen. None had a history of smoking, and none had bone-destructive lesions in the mandible. No subject had menstruated for at least 1 year.
Panoramic radiography were taken for all subjects with informed consent at the time of DXA measurements of the lumbar spine (L2-L4). All panoramic radiographs were obtained with a AZ-3000 (Asahi Co., Kyoto, Japan) at 12 mA and 15 s; kVp varied between 70 and 80. Screens of speed group 200 (HG-M, Fuji Photo film Co., Tokyo, Japan) and film (UR-2, Fuji Photo Film Co., Tokyo, Japan) were used. Appearance of the mandibular inferior cortex was bilaterally clear in the radiographs. All radiographs were digitalized with the resolution of 300 dpi using a flat-bed scanner (ES-8000, Epson, Japan).
When using the definition of the Japanese Society for Bone and Mineral research, 11 54 of the 100 women presented normal BMD (BMD more than 80% of Japanese young adult), 21 osteopenia (70-80%), and 25 osteoporosis (less than 70%) in the lumbar spine. The rate of women with osteoporosis in the lumbar spine in our study was similar to that (26%) in 1,033 postmenopausal women aged 50 years or older in the Adult Health Study (AHS) cohort in Japan. 12 Region of interest (RoI) was taken from four different areas of an image. Two RoIs were taken from the left hand side and the right hand side. The location of trabecular bone samples were between root of the tooth and cortical bone.
Area where sample was taken on panoramic radiographs is illustrated in Figure 1a. Figure 1b and Figure 1c show samples from left hand side, whereas Figure 2d and 2e are samples from right hand side. In this experiment, each sample size which marked with white box is 128 × 128 pixels.
Multiscale line operator is one of line detection algorithm used for detecting linear structure on mammographic image together with other line detection methods. 13 Comparing with other methods, the line operator algorithm was proved to give good result from signal to noise aspect, line width accuracy, and localization. In the early implementation of multiscale line operator algorithm, the algorithm has been used to detect asbestos fiber. 14 This algorithm was applied for detecting linear structure of iris blood vessel. 15 Multiscale line operator requires parameters of angle (q) and length (M). 16 Angle controls the number and size of analyzed rotation. Angle size per rotation was sum of the current and previous angle size until the limit 180 degree. Length parameter was needed to make moving window with length M. We used 12 rotations with angels of 0, 15, 30, 45, 60, 75, 90, 105, 120, 135, 150, and 165 degree. Moving window size of 5 × 5 was applied to each pixel for analysis it with 24 neighbor pixels around. Given a region of interest image, at each pixel (x, y), multiscale line operator algorithm measures the line strength S(x, y) by calculating the contribution of the foreground minus the contribution of the background. For each angle q, the foreground mask has a line of length M and width one pixel, oriented at the angle q. The foreground value, F(x, y, q), was the sum of pixel values multiplied by the corresponding foreground mask values. Similarly, the background mask was a rectangle of size M × M, oriented at the angle q. The background value, B(x, y, q), was the sum of the image pixels multiplied by the corresponding background mask elements. The line-strength image value, S(x, y), was calculated by: Each segment of the line structure has different width. To estimate the problem, we require multiscale analysis method. In this paper, we use Gaussian pyramid to solve multiscale analysis problem. Gaussian pyramid is generated by first smoothing the image with an appropriate smoothing filter and then subsampling the smoothed image as many as desired levels. This produces a set of gradually more smoothed images. Nevertheless, the more smooth an image, the less sampling density. If we illustrate it graphically, this multiscale representation will look like a pyramid, from which the name has been obtained. Images produced at the lower levels of the pyramid have higher resolutions, whereas those produced at the higher levels have lower resolutions.
The levels of the pyramid obtained iteratively as follows. For 0 < l < N: However it was convenient to refer to this process as a standard REDUCE operation and simply write: where w (m, n) was "generating kernel". The structure of generating kernel is [0.25 -a/2, 0.25, a, 0.25, 0.25 -a/2] and a = 0.4. The equivalent weighting functions were particularly Gaussian-like when a = 0.4, when a = 0.5 the shape was triangular, when a = 0.3 it was flatter and broader than a Gaussian. With a = 0.6 the central positive mode was sharply peaked and was flanked by small negative loobs.
After the line strength structure of the trabecular bone was detected, those images were transformed into binary images. This method changes the background into black with zero value and changed the image object (foreground) into white with one value.
Measurement of line strength in trabecular bone was applied to the binary image at each pixel based on its existence on the trabecular bone with prioritizes line segment which has similar orientation with the root of tooth. The total value of line strength was the mean of line strength in four samples on every panoramic radiographs. The number of the bone segment was used to decide whether the patient was affected osteoporosis or normal.

results
Sample image from panoramic radiographs was processed using multiscale line operator algorithm and Gaussian pyramid. Figure 2a, 2b, 2c, and 2d show result images from detection of trabecular bone linear strength structure applied to images in figure 1b, 1c, 1d, and 1e, respectively. Output for image in figure 1b and 1c from line strength structure detection using orientation 0, 15, and 30 degrees is shown in Figure 2a and 2b. Figure 2c and 2d show images from line strength detection using orientations 330, 345, and 360 degrees for image shown in Figure 1d and 1e, respectively.
After processing with multiscale line operator algorithm and Gaussian pyramid, sample of trabecular bone image was converted to binary image. The total value of line strength from four black and white samples on every panoramic radiographs was added. After that, mean value of line strength on one panoramic radiographs was calculated. From 100 panoramic radiographs used in this experiment, line strengths were compared with threshold.
The cut off threshold of trabecular bone line strength was selected at 3450 by 92% sensitivity. The result of line strength and its correlation with osteoporosis status was shown in table 1. The average pixel number of black and white image was used to diagnosis osteoporosis by comparing the number with threshold. This number was preferably 3400-3500, and more preferably 3425-3475, and 3450 was the best threshold value for the RoI size of 128 × 128. The number of subjects identified by mandibular trabecular density less than and equal to the threshold were 10 and 23 for normal and osteoporotic subjects. There are 62 subjects for trabecular density higher that the treshold and 2 subjects for normal and osteoporotic, respectively.
Image that has a wide average of line strength of segment structure of trabecular bone means that the image was not affected by osteoporosis. Image that has a small average of line strength of segment structure of trabecular bone means that the image was potentially affected by osteoporosis. Diagnostic efficacy of manual and computeraided automatic measurements is presented in Table 2. This method achieved the accuracy of 88%, sensitivity of 92%, and specificity of 86%. The correlation between BMD and trabecular bone line strength was 52%.

discussion
Various diagnosis and treatment for osteoporosis has been discussed and some of them has been implemented. 17 Even though there have been several studies regarding osteoporosis on postmenopausal women in Indonesia, 18 however this study is the first demonstration comparing correlations between skeletal BMD and trabecular bone line strength measured by a computer-aided system on digitized panoramic radiographs for postmenopausal women. Line strength measurement has been used for robust classification of anatomical types (vessels, spicules, ducts, etc) in a mammogram. 13 In other study about trabecular bone for osteoporosis detection, it shows 92% sensitivity and 96% specificity. 8 We have the same sensitivity but lower specificity. In our previous study, 19 sensitivity and specificity for computeraided system were about 88.0% and about 58.7%, respectively. Although it used cortical bone on mandible as subject, similar result can be got by using trabecular bone. With this image analysis system, line strength of trabecular bone can be measured and compared with radiographs as the input file for the system. Using conventional DEXA scanner cost about IDR900.000,-for assessment on lumbar spine and femoral neck. While this system only need to have a panoramic radiograph and the assessment system which cost about IDR100.000,and IDR50.000,-respectively for each assessment. Panoramic radiograph can be easily taken in clinic and then scanned. Considering these advantages, it is very possible to implement our proposed system to examine Indonesian woman.
This computer-aided system, however, has some limitations. Dentists were asked to determine manually RoIs along trabecular bone area. Error tends to occur with this determination due to the existence of root. Automatic determination of the RoIs would be necessary to maintain good reproducibility around the world. The robustness of this system would also be necessary to overcome this system limitation.
In conclusion, our approach of analyzing trabecular bone using panoramic radiographs has sensitivity and specificity of 92% and 86.7%, respectively. Thus, we suggest that the computer-aided diagnosis system may be useful for detecting osteoporosis.