Segmentation and Classification of Jaw Bone CT images using Curvelet based Texture features

The evaluation of jaw bone trabecular structure and quality could be useful for characterization and response of the bone for dental implants. Current clinical methods for assessment of bone quality at the implant sites largely depend on assessing bone mineral density using Dual energy X-ray absorptionometry. However, this does not provide any information about bone structure which is considered to be an equally important factor in assessing bone quality. This paper presents a novel approach for computer analysis of trabecular (or cancellous) bone structure using multiresolution based texture analysis to evaluate changes taking place in the architecture of bone with age and gender. The findings are compared with Hounsfield Units measured from the CT machine at different sites, which is a standard reference. Fifty patients were subjected to clinical CT to obtain the CT number and texture based architectural parameters respectively. In each site texture features were extracted using gray level co-occurrence matrices (GLCM), Run length matrices, Histogram and curvelet based statistical & co occurrence analysis. A very difficult problem in classification techniques is the choice of features to distinguish between classes. However the performance of any classifier is not optimized when all features are used. The feature optimization problem is addressed using Principle component analysis in terms of the best recognition rate and the optimal number of features. Testing this on a series of 120 image sections of trabecular bone with normal, partial and total edentulous patients correctly classified over 90% of the porous bone group with an overall accuracy of 87.8%–95.2%.The results shows that by using the Classification & Regression Tree approach the combination of the features from gray level and Ist order statistics achieved overall classification accuracy in the range of 87.890.24%. Features selected from the curvelet based co occurrence matrix performed better with overall classification accuracy of 92.89%.In order to increase the success rate the classification is done using the combination of curvelet statistical features and curvelet co occurrence features as feature vector and using this, a mean success rate of 95.2% is obtained.


Introduction
Since early times of implantation era preoperative studies includes incision of gingiva in order to get a view of the bone surface.Preoperative studies required because a jaw bone must offer proper quality and adequate quantity of the bone.The overall dental implant success rate is considered to be influenced by both the volume (quantity) and density (quality) of available bone for implant placement.Successful implants require good bone and plenty of it.The thickness of the bone and height of the jaw bone available is easily measured on CT scans.Quality of bone is one of the variables that, unfortunately, cannot be accurately determined prior to the placement of the implant.Bone quality, as described by Lekhom and Zarb [1], is of major importance for the success of an implant placement.For preoperative planning bone quality has been categorized into five classes that basically describe the relation of cortical and cancellous bone in a specified region of the jaw [Table 1].The amount of cortical bone is responsible for the primary stability of the implant, where as cancellous bone is responsible for long term stability.Hence the characterization of cancellous bone architecture is of considerable interest to evaluate quality of the bone.At present, only the histomorphometry techniques [2] allows determination of cancellous bone architecture of a patient.However, it is an invasive technique, as a biopsy of around 1 cm in diameter is taken from the patient.As cancellous bone has anisotropic mechanical properties, textural parameters as related to trabeculae orientation may give clinically valuable information.The proposed method focuses on image based bone density classification that utilizes image gray scale values and multiresolution texture based image features.This work develops an algorithm that is able to support surgeon's decisions based on analysis of CT images of jaw bones.Dental CT is frequently used as a non-invasive in vivo measurement tool for cancellous bone density, the denser the bone, the higher amount of X-ray energy is attenuated, resulting in varying intensities in resultant CT data sets [3].CT image inferred bone density and mechanical property relationships are important for osteoporosis research, computational modeling of patient specific bony anatomical structures, surgical planning and tissue engineering.
The micro architecture of open celled cancellous bone is highly anisotropic, and varies with age, medical condition and gender.This variance is present as density measurement which is based on the amount of bone tissue with in a cancellous bone specimen.In conjunction with bone density readings, a qualitative measurement of structural parameters may improve our ability to estimate bone strength in better way [4].Frequency analysis obtains the image features relating to its structure which is useful in predicting bone properties.Multiresolution analysis allows for the preservation of an image according to certain levels of resolution or blurring.Broadly speaking, multiresolution analysis allow for the zooming in and out of the under lying texture structure.The wavelet based multiresolution analysis has been used in feature extraction since it has the ability of localization in both time domain and frequency domain, and can also focus in any detail of the object [5,6].Therefore the texture extraction is not effected by the size of the pixel neighborhood.By decomposing the image into a series of high pass and low pass bands, the wavelet transform extracts directional details that capture horizontal, vertical and diagonal activity.However, these three linear, directions are limiting and might not capture enough directional information in noisy images, such as medical CT scans, which do not have strong horizontal, vertical and diagonal directional elements.
Ridgelets, like wavelets, provide multiresoltion texture information; however they capture structural information of an image, based on multiple radial directions in the frequency domain.The authors show in [7] that, the multi directional capabilities of the ridgelet transform provide better texture discrimination than its wavelet counter part.However, one of the limitations of this approach is the fact that ridgelets are most effective in detecting linear radial structures, which are also not dominant in medical images.A recent extension of ridgelet is the curvelet transform [8,9].Curvelets are proven to be particularly effective at detecting image activity along curves instead of radial directions.Curvelets also capture structural information along multiple scales, location and orientation.Instead of capturing structural information along radial lines, the curvelet transform captures this structural activity along radial "wedges" in the frequency domain [10,11].In view of that, curvelet transform is firstly proposed to highlight the curved singularities in images for structure recognition.

Curvelet transform
Candes and Donoho in 2000 introduced the curvelet transform [8,[10][11][12].The continuous curvelet transform can be defined by a pair of windows W(r) (a radial window) and V(t) (an angular window), with variables W as a frequency-domain variable, and r and θ as polar coordinates in the frequency-domain.
A polar 'wedge' represented by U j is supported by W and V, the radial and angular windows.U j is defined in the Fourier domain by U j (r,θ) = 2-3/4 W(2 -j r)V(2 ∟j/2 θ)/2π (3) The curvelet transform can be defined as a function of x = (x1, x2) at scale 2 -j orientation θ l , and position x k (j,l) by where R θ is the rotation in radians.Fig1 illustrates the polar 'wedges' represented by U j .Further details are presented in [].

Texture analysis
Traditionally, texture features have been calculated using a variety of statistical, structural and spectral techniques including co-occurrence matrices [13], run length statistics [14], spectral measures, fractal dimensions, statistical moments, and multi resolution techniques such as wavelets [6], ridgelets [7] and curvelets [10,11].Cooccurrence matrices are often used in texture analysis, since they are able to capture the spatial dependence of grey level values within an image.Common statistical measures used in texture analysis are mean, standard deviation, energy, entropy, contrast, homogeneity , variance, correlation, maximum probability, summean, cluster tendency and inverse difference moment [15].In this paper, the curvelet transform is applied on a 3D data set of DICOM images of 50 patients.Curvelet statistical features, such as mean and standard deviation obtained from the first level of sub bands and curvelet cooccurrence features are extracted from each of the curvelet sub bands from higher level wedges.The structural activity extracted from the curvelet transform of the image can be analyzed statistically to generate curvelet based statistical and cooccurrence texture features.Correlation coefficient is calculated to find the relationship between feature change and percentage calcium loss in different directions and to determine which features undergo changes most closely corresponding to bone mineral loss.

Methodology
The algorithm proposed in this article consists of four main steps: Segmentation of regions of interest, application of the discrete curvelet transform, extraction of texture features, statistical analysis between HU and texture features based on gender and age and creation of a classifier.The following sections describe the data set, region of interest calculations, texture extraction methods and classifier performance.

Data set
From June 2006 to October 2008, 50 patient's age: 30 ~ 71 years with a mean of 50 years; 26 males and 24 females received the spiral CT examination.All patients needed dental implantations either in maxillae or mandible in selective locations.Single, partial, or total edentulous patients were included.Axial, panoramic and cross sectional images are acquired at each bone site with multi slice spiral CT scans (120 kv and 300 mAs, with a slice thickness of 0.625mm, pitch 0.4mm, scan time 750msi GE medical systems).The 3D DICOM image data consists of 190 consecutive 2D slices, each slice being 512x512 pixels in size and having 16 bit gray level resolution.The mean bone density of the implant recipient site was measured by using software incorporated in CT machine (GE).The bone density measurements were recorded in Hounsfield Units (HU).The scan was initially utilized to assess bone quality subjectively based on Lekholm and Zerb (1985) classification at anterior and posterior edentulous sites by rating the distribution of cortical and cancellous bone.Density of cancellous bone in Hounsfied units [HU] was determined by averaging the readings of multiple slices within respective sites.The classification of bone quality of each site and the CT analysis of cancellous bone densities are listed in Table1 and Table3.The density in HU for the cancellous bone sites varied widely, especially within the mandible, irrespective of the amount of cancellous bone.Curvelets like wavelets are extremely sensitive to contrast in the gray level intensity, the segmented images need further processing.In order to effectively use curvelet based texture descriptors, it is necessary to eliminate all background pixels to avoid mistaking the edge between the artificial background and the tissue as a texture feature.Each slice was therefore further cropped, and only square 32x32 sub-image fully contained in the interior of the segmented area are generated [see Fig 2 and Fig 3].This size was chosen since the digital curvelet requires a 2n square image [9,10].

Feature extraction
Once the bone images were pre-processed the texture feature vectors used in the algorithm are derived from the Gray level run length matrix, image histogram and Discrete Curvelet Transform, introduced by Candes and Donono in [9].Bone trabeculae appears in the image as an array of connected pixels having the same gray level range, which corresponds to the definition of a run.However, a 65,256

Results and Discussion
The discriminating powers of the feature vectors, the feature variations in male and female at various implant recipient sites and comparison of curvelet based descriptors with co-occurrence and run length based techniques is described in the following section.The mean and standard deviation for each of the texture parameters from the 50 bone images (26 Male, 24 Female) are presented in Table 2.The CT slice of mandible and maxilla was shown in Figure 4.
Three possible changes in image texture were observed based on the parameters of Table 2: First, as de calcification progress there could be an even reduction of intensity along the profiles; that is, no textural change would be observed apart from a mean decrease in intensity.Another is uneven or patchy bone demineralization with an initially dramatic increase in texture roughness at points along each profile.Finally, there could be an uneven reduction in profile intensity but with a smoothing of the image texture.Examination of changes in a gray level statistic matrix during demineralization will help to clarify these findings.For a typical image profile, non zero matrix elements were dispersed along or near its

40
A visual representation of these values with 95% confidence intervals is shown in Figure 5 and the polynomial regressions are shown in Figure 6.The texture parameters derived by the runlength matrix, curvelet based statistical and co occurrence features on jaw bone CT images constitute a computer based assessment of bone quality.The minimal generalized calcium loss detected by these features is still a substantial improvement over the estimated 30-60% decalcification that must exist before a radiologist is certain of bone loss using conventional radio graphic interpretation.Bone in the maxilla should be loss concurrently with or even earlier than other axial bone.Texture features extracted based on curvelet transform and distribution of these features in various directions assist in characterizing alveolar bone.In this study ten out of twenty four texture parameters could predict changes in trabecular network.The alterations in the trabecular network with respect to advancing age were also quantified by the texture analysis.This noninvasive analysis may provide information about the trabecular architecture that is independent of bone mineral density but depends on age.One can conclude that by measuring changes in statistical texture parameters it is possible to monitor changes in calcium contents and internal structure of the bone.The results of this study indicate that curvelet based texture analysis is an effective tool that could quantify bone architecture as well as bone density in two dimensional bone slices.The major advantage of using a multi-resolution approach in this application is that it examines the directionality of the texture components at all scales (resolutions) simultaneously and independently.This method can be used by the implant surgeons to know about the bone quality at the implant sites pre operatively which can avoid the cost and trauma undergo by the patients.

Figure 2 :Fig 3 :
Figure 2: Segmented bone image with region of interest.

Table 1 :
[1]el image (16 bit image slice) yields a huge number of runs having a small length which are not representative of the bone architecture, as the pixels in a trabecula do not have the same gray level.If the image is reduced to 64 gray levels, the trabeculae then appear in the same gray level range, and thus have change to form a run.So the original 16 bit CT images were quantized into 64 gray levels using linear mapping.From the gray level run length matrix features such as GLNU, RLNU, RP etc are calculated and stored as feature vector F1.Classification based on CT determination of Bone density[1]

Table 2 :
Gender wise significant texture parameters with varying age group

Table 3 :
Bone quality classification vs. measure of cancellous bone density in HU for all sites

Table 4
shows a comparison of the performance of five feature vectors using CART approach.The overall classification accuracy result achieved was 95.2 % (specificity 90.11%, sensitivity 92.13%)

Table 4 :
Classification performance for all the feature vectors *Data are expressed as percentage.PPV-positive predictive value; NPV-negative predictive value.