New Segmentation Approach to Extract Human Mandible Bones Based on Actual Computed Tomography Data

In this paper, a new approach for segmenting different anatomical reg ions in dental Computed Tomography (CT) studies is presented. The approach consists of three steps: Hounsfield unit's threshold (HU) based on gray-level segmentation, multi-object with textu re extraction and anatomical regions identification. First, a HU threshold window sets to separate between different regions upon their gray-level values; second, a set of objects are generated by and texture descriptors are calculated for selected windows from the image data sample. Finally, identification of different anatomical regions set for mandib le bones cortical and cancellous. It is expected that the proposed approach will also help automate different semi-automatic segmentation techniques by providing initial boundary points for deformable models or seed points for split and merge segmentation algorithms. Preliminary results obtained for dental CT studies of human-mandib le are presented.


Introduction
Image scanners devices such as computed tomography (CT), magnetic resonance imaging (MRI) or positron emission tomography (PET) are nowadays a standard instrument for diagnosis. Among these devices, CT-scanners are today widely used at radiotherapy departments all over the world it has several advantages. The main advantages of a CT-scanner are to obtain physical informat ion, like patient anatomy, size, shape, and in homogeneities; the other is to obtain the electron density fro m the different anatomical structures of the patient for the radiotherapy treatment planning [1]. Image segmentation is one of the primary steps in image analysis for object identificat ion. The main aim is to recognize ho mogeneous regions within an image as distinct and belonging to different objects, where the seg mentation process can be based on actual Dig ital Imag ing and Co mmunicat ions in Medicine DICOM data to find the maximu m ho mogeneity in grey levels within the regions identified.
A variety of techniques have been proposed for CT image segmentation methods. These methods can be classified into t h ree catego ries . Th e first categ o ry includes analyt ic methods, in wh ich seg mentat ion algo rith ms are t reated direct ly by cons idering so me measu re (e.g. Hounsfield Unite), wh ich by a priori knowledge is assumed to be the appropriate measure, Gabriella  extract the mandib le contour by h istogram equalization and thresholding and proposed another method by gradient vector flow snake parameters were optimized in order to achieve more accurate contours segmentation of nerve mandibular scans [2]- [4]. Paola Campadelli et al (2009) apply hierarchical g ray level based on framework to segment heart, bones, liver, kidneys, and spleen directly related to the Hounsfield units (HU) based on gray level techniques, learning techniques, model fitting techniques, probabilistic atlases, and level set anatomical knowledge to obtained three dimensional (3-D) binary image [5]. Typically this measure was incorporated into the original segmentation algorithm as well.
The second category includes supervised evaluation methods. In these methods, the results of a segmentation algorith m are co mpared to a "standard" reference image that is manually segmented beforehand, and the amount of discrepancy becomes the measure of seg mentation effectiveness. Gaivile Pileicikiene et al (2007) choose dead person for the tudy the precision of CT examination increased -oluntary and involuntary movements of research object were excluded (like breathing or muscle tonus movements), this make the segmentation more easier since the high resolution images captured by increasing the dose to get more than 1500 slice fro m the CT scanner [6], Ming Chen et al (2007) choose the same condition with another method by comparing between dead and alive person. The author used commercial software to segment the DICOM images based on threshold and region growing methods to extract the mandib le for a patient [7]. Zhan Liu et al (2007) used the same commercial software in his studies, but here the Temporo mandibular Jo int (TMJ) appear clearly in the 3D model geometry informat ion of the cortical bones, the cancellous bones and the teeth was exported fro m commercial software and this may accomplished by semi-automated method [8]. This is the most commonly used method for med ical segmentation. However, manually generating a reference image is a difficu lt, subjective, and time-consuming job, and generally cannot guarantee that one manually-generated segmentation image is better than another. Consequently, comparison using such reference images cannot ensure accurate evaluations.
The third category includes unsupervised evaluation. In these methods, the segmentation results are evaluated by judging the quality of the segmented image directly to evaluate some pre-defined criteria, such as the partitioning of foreground objects fro m the background, J.M. Reina et al (2006) present a 3D surface model o f mandib le based on morphological analysis with standard distances without case study and this near to the information theory. Since the shape of each organ is not consistent throughout all slices of a 3D med ical image and the g ray level intensities overlap considerably for soft tissues, texture is especially important in medical image segmentation because of its homogeneity within the same t issue and across different slices. Once textures have been calculated and their scalar values assigned to pixels, the pixels can be clustered or classified (when the tissues' labels are availab le) for the purpose of segmentation [9]. These evaluation measures are typically used with gray-level images and are not designed for general-purpose applications. However, modeling of biological t issues, such as bone related organs, is a difficu lt task because of their inherent inhomogeneous and anisotropic character, although inhomogeneous character is in some sense "directly" accessible via imag ing techniques. Unfortunately, this is not the case for the trajectories of anisotropic elasticity. Many investigators have evaluated 3-D models fro m Co mputed Tomography (CT) 2-D images. The results of these studies have been approximated for the conversion of images formats. Mechanical anisotropy means that the mechanical properties of the material are different when measured in different d irect ions in the same sample. For reasons of simplification, modulus of elasticity of the mandibular bones, was given constant values and considered isotropic to facilitate simulat ion. [10] The use of numerical methods such as finite element methods (FEM) has been adopted in solving complicated geometric problems, as it is very d ifficu lt to achieve an analytical solution. FEM is a technique for obtaining a solution to complex mechanical problems by d ividing the domain p roblem into a collection of much smaller and simp ler do mains (elements) where field variab les can be interpolated using shape functions [13]. In 1977, Weinstein [11] was the first to use FEM in imp lant dentistry. Subsequently, FEM was rapid ly applied in many aspects of implant dentistry. Atmaram and Mohammed [12]- [14] analyzed the stress distribution in a single tooth implant, to understand the effect of elastic parameters and geometry of the imp lant, imp lant length variation, and pseudo-periodont al ligament incorporation. Borchers and Reichart [14] performed a three-dimensional (3-D) FEM of an implant at different stages of bone interface develop ment. Cook, et al. [16] applied it in porous rooted dental imp lants. Meroueh, et al. [17] used it for an osseointegrated cylindrical implant. W illiams, et al. [18] carried it out on cantilevered prostheses on dental implants. Akpinar, et al. [19] simulated the combination of a natural tooth with an imp lant using FEM.
In the present study, a new approach for segmentation based on actual CT data are used to extract an accurate human mandible with TMJ and teeth to help for generating robust 3D surfaces and volu metric models. The proposed technique used the statistical texture method applying entropy transformat ion comb ined with informat ion captured fro m HU based on gray-level techniques, finally a level sets and geometric contour used to build the prototype geometry for the full-hu man mandib le.

Material and Methods
Dentate and edentulous patient' s mandible was scanned with SIEM ENS/ Esprit CT machine 120 KeV energy. The pixel size of the scanner is 512x512 p ixels, and 0.4 mm is equal to the distance between CT slice p lanes. Totally 400 images were obtained, raw data was in DICOM format.

A HU and Gray-level phase
The electron density is obtained from the CT-scanner via the so-called Hounsfield units (HU). These are defined as: where (μ) referred to the linear attenuation coefficient for the respective material co mpared with water. The linear attenuation coefficient depends on parameters such as electron density, atomic number and the beam quality of the CT-scanner. In the imp lementation of the Hounsfield scale in this study the Hounsfield scale stretches between HU = -1024 and HU =3071.A HU of -992 represents air outside the patient and a HU of larger than 2832 represents iron.
DICOM fo rmat files carry ing all information about CT scanned, different tissues segmented based on HU range by verifyed windows defin ing the tissues pixel positions, each slice has different HU range and this helps in applying the threshold technique to evaluate the difference in bone density for different bone types, where any noise maybe produced by the reflection rays on filling materials affects the HU range and the 3-D reconstruction of the mandible.can be reduced by used segmentation techniques based on selecting the HU values, H min and H max. Figure 1 shows CT slices and its histogram depends on HU range created by MATLA B program, this helps to extract bones fro m slices according to its HU range. HU min and HU max are reported at table 1, and the different tissues have different values of HU    A Multi-object reconstruction and texture phase The mu lti-object reconstruction helps to extract the mandib le bone fro m skull; that the model will have specific informat ion about mandible, proposed segmentation technique is used to extract it fro m the CT slices based on region growing technique, which is imp lemented with MATLAB program. Triangulation process method is used to generate the 2½-D models and the marching cubes for the final 3-D model; this is applied for all slices.
After the mu lti-object reconstruction process; the objects must be enveloped by texture material as shows in Fig. 3. Multi-objects regions enveloped with texture material to verify their prosperities, by evaluating the vertex (x,y,z) of model fro m origin (0,0,0) p ixels positions can be calculated. Intensity can be measured by calculating the pixels values mu ltip lied by pixels numbers (3). Preliminary intensity measurements are shows in table 2.
where P n equal to pixel position and P v equal to pixel value and i refer to pixel nu mber.

Results
There are three different models provided fro m this approach, Fig.4 represents models for the dentate patient, who used to validate the technique by comparing it with the provided model fro m the phantom software at the CT scanner, full sku ll with and without skin with different colors to show the different objects. Table 3 shows the comparison between Multi-object models with DICOM images and phantom software at the CT scanner, where the TMJ that hides behind the skull bone appears with Multi-object models, the artifacts noise from felling metals or imp lants and undesired soft tissues will be reduced.   Finally the cancellous bone pattern can be exported to any fin ite element program with STL format as shown in Fig. 7, where Abacus program are used to provide a finite element model for the trabicular pattern divided by tetrahedral elements.

Conclusions and Future Works
Many investigators have evaluated 3-D models fro m CT 2-D images. The results of these studies have been approximated fo r the conversion of images formats, this study provide a realistic 3-D model fro m 2-D segmented images scanned with CT scanner device based on actual CT data carried by DICOM files formats. Preliminary results are reported; pixels intensity values and pixels numbers are significant to verify the volu metric 3-D model of objects, this model can helps in dental diagnosing and to evaluate the orthogonal mechanical properties values especially for edentulous patients. The noises produced for filling cavity materials or dental imp lants are eliminated by combined classical segmentation techniques. Tetrahedral elements are suitable to divide the mandible model to fin ite elements for mechanical analysis studies like imp lant stress-strain; mu lti-objects technique separate mandible to objects for decreasing the number of nodes and elements to reduce the number of FEA equations; this will accelerate the idealization process at FEM models.