In the dental computer-aided design and computer-aided manufacturing (CAD/CAM) industry, volumetric computed tomography (CT) images and scan surfaces are most commonly used. However, the two data types are very different, because their measurement techniques fundamentally differ. The volume data contains intensity information of the internal organs of the human body, while the surface data contains only the visible surfaces, that is, the teeth and the gingiva. Because of their different features, volume data and surface data are used for different dental applications. However, there are many applications which require both volume data and surface data and for the accurate registration of the volume and surface data is necessary. To achieve this we propose a novel registration method of volume data and surface data.
1.2. Related Works
Volume data and surface data have different features and there are various dental CAD/CAM applications which use both volume and surface data. Therefore, the registration of volume data and surface data is necessary.
Before approaching the registration problem, the intrinsic errors of each data should be considered numerically. 3D dental scanners (Identica Blue, MEDIT Corp., Seongbuk-gu, Seoul, Korea) are accurate to 0.007 mm. However, considering the whole process of making the impression and the plaster model for measurement, the total intrinsic error of the surface data is around 0.06 mm in practice [11
]. On the other hand, the accuracy of CBCT (MercuRay, HITACHI, Chiyoda, Japan) is approximately 0.20 mm [13
]. Dental prostheses cannot be designed using CBCT volume data because of this relatively low accuracy. Thus, the intrinsic error of the scan-derived data is generally negligible and only that of the volume data is a cause for concern.
Usually the registration problem concerns the same types of data and that is the basic premise in 2D images or 3D data registration. However, the registration problem in this paper concerns different types of data, volume data and surface data. Data must be converted to identical data types before the registration process, and most previous works convert the volume data to surface data. This type of conversion process that extracts dental surface data from volume data is called segmentation. Surface data registration can be performed on the resulting segmented dental surface data. Generally, surface registration is done by the iterative closest point (ICP) method, which needs good initial conditions [14
] and is widely used in the dental CAD/CAM industry. The flow chart for the ICP method is shown in Figure 1
Although the established registration framework (ICP) is currently used clinically for dental applications, some drawbacks still exist; the requirement of human inputs and the metal artifact problem.
Human input is needed to set the initial positioning. Although there is much research on global registration, which obtains the initial conditions automatically for ICP, applying this algorithm to dental model registration is challenging because dental surface data suffers from ambiguity due to teeth shape characteristics [18
]. Segmentation also requires human input. The most commonly used segmentation methods are thresholding, region growing, and active contour methods such as a level set. Thresholding is the most straightforward and basic segmentation method and teeth volume data is segmented by giving lower and upper intensity values [19
]. The region growing method starts with a set of seed points and regions are grown based on the similarity of intensity [20
]. Level set segmentation is performed using 2D axial direction sliced images and stacked for a 3D segmentation result [22
]. To improve the result of segmentation, combining the above segmentation methods has also been studied [25
]. However, every segmentation method mentioned above needs a human input; lower and upper threshold values must be defined for thresholding segmentation, seed points must be defined for region growing segmentation, and initial contours must be defined for level set segmentation. In the established framework for the registration of volume and surface data, this human input stage for segmentation represents the most time-consuming step. Although some automatic tooth segmentation methods have been studied, each study has constraints and cannot be used in a wide variety of applications [26
]. Also, once the segmentation has been done, post processing, such as surface smoothing and island removing, is required. The result of segmentation may even differ from person to person.
Another drawback of the established registration procedure is the metal artifact problem [28
]. If a patient has a prosthetic appliance made of metal, the volume data is seriously affected by white saturation. The quality of the resulting segmented surface is also affected. Initial conditioning for ICP also becomes more difficult because non-artifact points on the segmented surface must be selected manually. Many studies have considered the metal artifact problem [29
] but they have resulted in a reduction rather than an elimination of the metal artifact effect so the problem remains unsolved.