We divided our workflow into three steps, each of which requires a different file format. Step 1 involves acquiring a 3D volume image of the patient as a DICOM images file. Step 2 entails segmenting the anatomical structure from surrounding structures and exported to virtual 3D model in STL file format. Segmentation of osseous structures and soft tissue is relatively easy. However, in many cases it is difficult to create an STL model for two reasons. One reason is that thin osseous structures (e.g. bone surrounding the nasal cavity, orbital floor), and narrow tissue gaps (e.g. upper and lower joint cavity between the temporal bone and the mandible) are not clearly reproduced in the STL model. Secondly, many artifacts (e.g. metal artifacts and/or beam-hardening from dental prostheses) reduce the readability of the images and prevent segmentation. Step 3 concerns 3D printing the physical 3D model, which requires use of “G-code” generation software to produce G-code as 3D printable data [13]. Each step of the entire process–segmentation of DICOM images, processing of STL data, generation of G-code data, and performance of the 3D printer itself–affects the accuracy of the final 3D model. Creating STL data is the most important operation in fabricating the 3D model.
Characteristics of DICOM segmentation and STL creation software
Appearances of the created STL models differed across software packages. Most notably, the cortical bone of the top and/or lateral pole of the mandibular condyle was thin, so the reproducibility of this part was different across all software packages. When “faithfully” fabricating according to this STL model, the steps would appear as holes (defects). Moreover, in some software packages, the surface of each STL model was rough (Fig. 7).
Although the ball STL model was created by MDCT scanning of a 10-mm-diameter bearing ball, all software packages rendered it expanded in all directions. The average ball length in all directions was 10.52 mm, but the length in the Z-axis direction was slightly longer than in the X- and Y-axis directions. This is likely because of differences in voxel size of DICOM images (X-, Y-, Z-axis direction lengths were 0.468, 0.468, and 0.500 mm, respectively), and may also have been affected by the partial volume effect that occurred on the border between the ball surface and the air. The diameter of the ball in the STL model was calculated from the mean value of the volume (605.23 ± 42.38 mm3) as 10.49 mm. The shape error for this entity was equivalent to the size of one voxel, and was reproduced by each software package.
It is difficult to quantitatively assess the STL segmentation performance of each software package independently. To solve this problem, we superimposed pairs of STL models (created with different software packages) on each other; the difference between each pair was visualized and measured as a shape error. Although differences between shapes of the created STL models were visible on the shape error image, no significant statistical differences were found across all mandible STL models. Figure 8 shows images captured by superimposition and visualization of S3D and MIT, which had the minimum shape error. Figure 9 shows images of MCS and VE3 having a maximum shape error. The reason the shape errors could be seen by the software packages, though only slightly, was that the binarization algorithms differ across software packages. Binarization, that is, creating an isosurface. The isosurface refers to the boundary surface of the target area formed by setting an appropriate threshold, and is generally approximated by a polyhedron as a patch model consisting of a set of fine triangles. The method of creating isosurface from volume data has been used in a wide range of fields as a useful tool such as 3D visualization of CT data and modeling of arbitrary shapes by implicit function expression. A number of methods have been proposed [14-16].
The shape error appeared because of differences in image processing near threshold values, such as the thin cortical bone or strongly curved surface. The color map of Figure 8 and 9 are colored as a green to yellow area, with mean distances of around 0.30 mm. This is smaller than one voxel size. Regarding the roughness of the surface of the STL model, it was thought that the influence of the unevenness was small. Therefore, it was considered that the shape error was not affected. It is difficult to judge the pass/fail of an error that differs depending on the software package obtained in this study because there is no correct answer. Considering the spatial resolution of MDCT, it can be assumed that this kind of error is acceptable in fabricating 3D models for clinical use in oral and maxillofacial surgery [17-19].
Reducing STL data size
STL data represent a 3D shape as a collection of small triangles. The number of triangles depends on the size, shape and internal structure of the object. More complex features and higher resolution lead to an increase in the number of triangles in the segmented STL data. Processing a large number of triangles draws heavily on the processing power of a PC; the calculation is time-consuming and can affect subsequent operations. Reduction in the number of triangles directly leads to a reduction in data size. However, a reduction in the number of triangles may also cause a morphological change [20]. Therefore, the mandible STL model was superimposed before and after the reduction in the number of triangles to evaluate the dimensional change, and the shape error was observed. To reduce the number of triangles to 200,000, i.e. the number of triangles recommended in the report [21], the “simplify data by specifying the number of triangles” function of PMV4 was used [22]. Figure 10 show before and after reduction of the number of triangles and the color map after the superimposition of the STL model with the largest volume and number of triangles (IN3; 1.24 million). As a result, although the surface of the STL model with the reduced number of triangles (200,000) was somewhat rough when displayed on the monitor, the resultant shape error of that STL model relative to the models with the largest and the mean numbers of triangles was almost 0 mm. It was clarified that data reduction of the mandible STL model of any software package could reduce the data size and did not affect the morphological change. Considering that the minimum laminating pitch of the FDM desktop 3D printer we use is 0.05 mm, this supports the inference that the recommended number of triangles was both necessary and sufficient for 3D printing.
Limitations and prospects
In the evaluation of the data size, the number of triangles, and the morphology of the created STL models, there was a problem that there was no gold standard value. Therefore, we solved it by performing multiple comparisons of all STL models. In this study, since only dry human mandible was used, segmentation operation with surrounding anatomical structures on a PC, such as soft tissue, was not performed. When performing 3D printing of a patient's DICOM data, segmentation of soft tissues and osseous structures is required. We have no manual measurement (e.g. measurement with a caliper) that is expected that measurement results will differ depending on the observer. Besides, the optical three-dimensional measurements that require verifying the accuracy of the measurement device itself in advance were not performed.
The shape errors are inevitable because of the spatial resolution limits of MDCT. However, when using 3D models in fields that require more detailed operations, such as microscopic surgery, other modality options should be considered, such as the use of limited cone-beam CT, which expected that produces a better high-definition STL model. In this study, a MDCT scanner was used to segment DICOM images to STL data under the condition of fixed voxel value binarization threshold. In addition to differences between patients, physics-based factors such as irradiation dose and other differences in MDCT models and scanning parameters may also affect the difficulty of creating STL models [23,24]. Although no segmentation in the true sense was performed in this study, in clinical use of 3D printing technology, setting a threshold for 3D printing requires medical knowledge, especially tomographic image anatomy, as well as knowledge of modalities of imaging principles. It seems necessary to understand the features of the software package for STL segments as well.
This study does not aim at ranking software packages. There are some differences between DICOM segmentation and STL creation depending on the software, so it is desirable to understand and use with those characteristics.