In oral medicine, a completely automated deep learning technique is gaining traction [30]. Recognition and segmentation of structures in panoramic radiographs, cephalometric radiographs, periapical radiographs, and CBCT images improve treatment efficiency, liberate the workforce from repetitive tasks, and create possibilities for instant assessment and massive workload [31]. Optical impressions can now be easily acquired, saved, and collected thanks to the rapid development of scanners, which aids orthodontic treatment and research. Its acquisition process required fewer invasive operations and enables for widespread adoption. It can also be used in conjunction with other digital approaches and serves as a critical data source for computer-aided design/computer-aided manufacturing (CAD/CAM) systems [32]. However, according to our limited understanding, artificial intelligence-driven tools for tooth surface and bracket separation and feature extraction from 3D scanning images have not yet been well introduced to the orthodontic field.
As computerization extensively develops in orthodontics, teeth and appliance recognition is the most fundamental process in CAD/CAM systems [22]. It also lays the groundwork for future instant and large-scale clinical evaluation. Recently, various segmentation frameworks have been developed and deployed in medical image recognition. However, most standard algorithms struggle to deal with irregular input data, making it difficult to recognize genuine teeth.. Thus, we use the pointnet neural network (PNN) to deal with the 3D scanned images for its high accuracy, invariance, and selectivity in recognizing and extracting irregular 3D image features [33]. The PNN-based network is more versatile and effective for tooth surface and bracket separation than traditional machine learning algorithms. It relies less on manual feature engineering, which is critical because mesh features can be abstract and practically hard to capture. While trained features are still mostly abstract and incomprehensible to human eyes, they are often not the primary concern for real-world tasks related to medical images or meshes.
The automated digital tool for tooth surface and bracket separation and feature extraction performs reasonably well on unseen data (new teeth and new bracket brand in dataset B). It recognized all bracket points, and the segmentation network also labeled all points as teeth for models of teeth without brackets (data not shown). It is partly because the segmentation boundary is relatively simple in nature, and point normal is a powerful feature in this situation. Nonetheless, the separation task in dataset B did not perform as well as it did in dataset A, owing to the network's lack of knowledge about the original teeth and brackets. Further enlargement of the training dataset can enhance the universality of the algorithm. The average time of tooth surface and bracket separation is about 2.9 ms per tooth and about 81 ms per case. Data preprocessing, including normal computation and point sampling, was executed by Python (with limited help from Numpy), and it took around 928 ms per tooth. A geometric processing library, such as VTK could reduce a great amount of time.
Aside from orthodontic bracket position assessment, the tool can be applied in treatment plan modification. Specifically, when patients request to switch to a clear aligner treatment in the middle of fixed orthodontic treatment for aesthetic concerns, this tool allows clinicians greater ease and flexibility in treatment designing and deciding when to change the appliance and whether to employ the segmental arch technique. Currently, it usually takes at least half an hour for an operator to manufacture a retainer manually or with CAD/CAM [34]. With this tool, instant retainer manufacturing will be possible. If clinicians perform oral scanning in advance, patients can wear retainers immediately after brackets are removed. Relapse is prevented and waiting time is reduced with the instant manufactured retainer [2].
In this article, we presented the most common application scenario, the bracket position assessment. Herein, we did not seek to understand which factors influence bonding accuracy; instead, we applied the tool in the evaluation scenario and built a working model. The tool can perform large-scale instant evaluation because of its high efficiency and independence from the original tooth and bracket data. The automation and low human interference of this tool allow it to be used as a standardized evaluation tool to assess and compare newly developed bonding systems, which saved a significant amount of time and made large-scale and instant evaluation possible [35]. It is preliminary work and potentially new. Nevertheless, it is a fundamental leap forward that eliminates repeated model scanning (or photography) and superimposing. In parallel, the result of the evaluation can be exhibited in two ways, both distribution patterns and deviation analysis linearly and angularly of the bracket position. This evaluation system and the way the results are shown could be used for immediate feedback and analysis in clinical practice and the training of residents.
The standards of idea bracket positions vary, and each has its own merits and flaws [36–40]. In this evaluation system, the midpoint of the facial axis of the clinical crown was advocated as the ideal bracket position since it is an acknowledged landmark in pre-adjusted appliances. Thus, it is suitable for this newly established system [38, 41]. The identification of FA was digitized, which essentially eased the workforce and eliminated human interference (i.e., different observation angles, personal preferences, etc.) [1, 5, 42–44]. However, flaws occasionally appear when teeth are not fully exposed, or when there is gingival recession or swelling. Further revisions will provide answers for the above-mentioned exceptional scenarios with the expansion of boned tooth data, virtual setup, and feedback from clinical practice [30, 31, 45].
It is worth noting that a great linear consistency can be observed in bracket positions on lateral incisors, which might benefit from their square-like shapes and flat labial surfaces [1]. The angular deviations are the smallest in the lower central incisors, probably due to their narrow and symmetrical shapes. Upper canines display the worst angular accuracy due to abrasion and uneven shape. Therefore, the direct bonding quality might strongly relate to the tooth morphology, which is in line with Hodge et al. [46]. Each orthodontist has an own preference, which may be related to training and treatment uncertainties [44].