The Diagnostic Advantage of a CBCT-Derived Segmented STL Rendition of the Teeth and Jaws Using an AI Algorithm

We can expect to see more and more applications of AI in our daily lives and in the orthodontic profession. AI will certainly affect the efficiency, accuracy, and quality of our orthodontic results. The application described in this article is another step forward in using AI to improve our treatment plans, treatment outcomes, and patient and parent communication.

MSL

The Diagnostic Advantage of a CBCT-Derived Segmented STL Rendition of the Teeth and Jaws Using an AI Algorithm

An accurate diagnosis is the key to a successful orthodontic treatment plan and, consequently, a successful treatment outcome. In today’s profession, this cannot be achieved without appropriate imaging.¹

The goal of imaging in medicine and dentistry, including orthodontics, is to display the patient’s “anatomic truth.” For many years, two-dimensional radiographs were used to evaluate the patient’s 3D status.² The use of 2D x-rays to analyze 3D structures has well-known limitations, however, including magnification, geometric distortion, superimposition of adjacent structures, projective displacement, rotational errors, and linear projective transformation.^3,4

Since its introduction to dentistry in 1998, CBCT has changed the diagnostic approach in both dentistry and orthodontics, providing accurate high-resolution 3D imaging from compact and affordable equipment. In addition, it exposes the patient to a significantly lower radiation dose compared with conventional CT.^5,6

Digital Process

CBCT is best described as “volumetric imaging.” An x-ray source and detector rotate in a cylindrical or spherical path around the center of the region of interest, with a single rotational sweep of 180-720° creating the field of view. Several hundred 2D images are generated in a procedure that takes 10-40 seconds. These images are processed by the computer to create rectangular cubes called voxels, which are .076-.4mm in size and equal in width, thickness, and height. Each voxel represents a shade of gray, depending on its x-ray absorption. Typically, one scan contains more than 100 million voxels. After the creation of a volumetric database from these voxels, different cross-sections can be produced in all planes of space (sagittal, coronal, and axial), in addition to curved cross-sections (such as panoramic images) and volume renderings (such as cephalometric images).^6-9

The 3D medical imaging data produced by the CBCT are stored and transmitted in the form of DICOM files. These files contain the raw data that become the basis for further manipulation.¹⁰ The DICOM file can display a series of 2D cross-sections in three planes of space—the most common rendition presented to the examining doctor. The DICOM file can also be segmented, allowing teeth to be removed or hidden so that underlying structures can be exposed. Additionally, the DICOM file can generate a 3D rotating picture of the field of view, including crowns, roots, and bone. The process of distinguishing root from bone is problematic, however, because the two structures are similar in density. Proprietary software programs are available to aid in this procedure, but they are labor-intensive, requiring substantial operator experience and time.¹¹ Even with the required expertise, the full length and anatomy of roots are often compromised or inaccurate. The pathway of the mandibular nerve can be drawn only manually, after examining a large number of 2D slices of the canal track.

To create a user-friendly 3D landscape of the DICOM data, the file should be converted to STL format. STL files, which are used for 3D printing of medical and industrial products, can be viewed on any computer with a standard operating system such as Microsoft Windows.^* Automatic STL conversion and segmentation can also be performed from a CBCT DICOM file of the teeth and jaws using an artificial intelligence (AI) program such as CephX,^** Anatomage,^*** or Imaging Plus.^†

AI capabilities have improved significantly since the 1980s and are now becoming important constituents of many medical applications. AI excels at recognizing complex patterns in imaging data, allowing accurate segmentation. In dental radiology, AI can identify the different tissues in a 3D DICOM package obtained from a CBCT scan and automatically convert it to a segmented STL image within minutes.¹² In orthodontic diagnosis, since all objects are segmented, the user can see different aspects and easily visualize various “what if” scenarios. For example, “What would the anatomy look like if a tooth were removed?” “Is there a hidden pathology beneath one of the structures?” “What is the spatial relationship between objects?” “Could the 3D interactive model offer better understanding simply by delivering the same information in a new package?”

These AI capabilities are based on deep learning mechanisms,^12-14 whereby input is received and the output is learned based on a large, labeled data set that is examined by the AI algorithm at the input level. The data set then undergoes more precise labeling,¹⁴ and the AI algorithm is taught to recognize it. Finally, the algorithm’s recognition capabilities are fine-tuned by human intervention (“teaching”) until optimal automatic recognition levels are achieved.

A 3D convolutional neural network is the most common machine-learning method of image recognition and classification.^12-14 Besides its uses in medicine and dentistry, it can be applied to such tasks as identification of objects, faces, and traffic signs and for computer vision in self-driving cars. Convolutional neural networks, like neural networks, are made up of perceptrons with learnable weights and biases. Each perceptron receives several inputs, weighs them, and passes the weighted sum through an activation function; it responds with an output, which, in turn, may serve as an input for another perceptron, and so on. In this manner, convolutional neural networks can mimic the decision-making tasks performed by our own central nervous systems.

The following two cases, both treated in the Tel Aviv University Department of Orthodontics, demonstrate how the AI capabilities of DICOM-to-STL conversion can have a significant effect on orthodontic treatment planning.

Case 1

A 13-year-old male presented with a mild dental Class II malocclusion, a bilateral crossbite (upper right first premolar through second molar to lower right first premolar through second molar, and upper left second premolar to lower left second premolar), and a functional shift of the mandible to the right (Fig. 1).

Clinical examination indicated a Class I skeletal pattern, maxillary constriction, and 9mm of crowding in the upper arch, with a lack of space for the canines; the lower arch showed moderate crowding (5mm). The left maxillary canine had partially erupted in an ectopic position, buccal to the upper left lateral incisor. The upper lateral incisors were buccodistally tipped, and the upper right canine was not palpable buccally or palatally.

The initial evaluation would likely have led to a nonextraction treatment plan including maxillary expansion and some form of arch development to resolve the crowding in both arches and gain enough space for the canines to erupt. There was, however, a radiographic suggestion of root resorption involving both upper lateral incisors. Because the panoramic and periapical x-rays could not clearly confirm the position of the upper right canine nor the integrity of the upper lateral incisor roots, the patient was sent for a CBCT scan.

The DICOM file of the scan was uploaded to the CephX web viewer for automated AI conversion to STL. A 3D segmented model of the teeth and jaws could then be viewed, and the canines and central incisors could be virtually extracted for better visualization of the lateral incisors (Fig. 2).

The STL web viewer of Case 1, which uses the proprietary CephX STL AI algorithm, is available at https://bit.ly/2Yjum5S. (This system is compatible with any smartphone with sufficient memory, but it was taken from a broad field of view and may take close to a minute to download.) To see how the 3D manipulation provides a vastly improved view of the maxillary lateral incisors, select Show/Hide under the View menu on the upper toolbar, then click on any relevant teeth and bony structures to toggle between shown and hidden. Close the Show/Hide menu by clicking Done.

Case 2

A 12-year-old female presented with a submerged lower right second deciduous molar and an impacted permanent lower right second premolar (Fig. 4A).

She was referred by a general dentist after complaining of pain in the lower right region, but she had no prior medical history. Clinical evaluation found a Class I skeletal pattern, a dental deviation of the lower midline to the right, and a posterior open bite on the right due to submergence and probable ankylosis of the lower right second deciduous molar. Neither the second decidous molar nor the lower right second premolar was visible clinically.

Radiographic examination revealed the positions of the submerged deciduous molar and the impacted permanent premolar, indicating their proximity to the inferior alveolar nerve canal and the lower border of the mandible, as well as the second premolar’s position relative to the adjacent first molar. A CBCT was requested for further investigation of this anatomical proximity and for surgical access planning (Fig. 4B.

An STL file was automatically converted from the DICOM file using CephX AI algorithms. The segmented model was virtually manipulated for better depiction of the 3D relationships between the objects (Fig. 5). Considering the proximity of the lower right second deciduous molar to the inferior alveolar nerve and the root of the first molar, it was evident that to avoid damaging those structures, the first vector of force on the lower right second premolar should be lingual and mesial. The STL model also demonstrated a severe mesio­lingual rotation of the impacted lower right second premolar, which was not apparent in conventional imaging.

Use the link https://bit.ly/3bHczd8 for 3D manipulation of the STL segmented model with the web viewer, following the same procedure as with Case 1.

Discussion

These cases substantiate the importance of 3D visualization in making an accurate diagnosis. Segmented STL models were instantly converted from standard DICOM data using an AI program.