Artificial intelligence applications in implant dentistry

In recent years, the evolution of modern implant dentistry and imaging techniques have been combined to create a new dental implantology protocol: guided implant surgery. With this innovative technique, a cone-beam computed tomography (CBCT) scan of the patient is taken to create 3D images of the oral structures. Afterwards, a specific software is used to design a surgical template, to establish the optimal placement (slope and depth) of dental implants. Although guided surgery ensures shorter surgical time and reduces risk of injuring important anatomical structures (i.e., nerves, bones, adjacent teeth, etc.), this technique is not commonly used as it requires longer preoperative planning time, normally performed by clinicians. The main aims of this study conducted at the company PrimoLab S.r.l. are to reduce the preoperative planning time by identifying the most critical planning steps from a time-consuming point of view, and to validate an alternative protocol of guided implant surgery, in which surgical plannings are carried out by biomedical engineers and templates are 3D-printed once clinician checks the virtual designs. Achieving these goals required operating at different levels. First of all, a time analysis has been performed with the Guided Surgery Team of PrimoLab, considering 200 cases of surgery; the analysis revealed that data manipulation’s step is crucial since it represents around 70% of the total preoperative planning time. Furthermore, around 85% of the total cases were categorized as belonging to mandibular planning; this can be justified considering that mandible is subjected to higher mechanical forces and consequently has a higher rate of healing than maxilla. Second, deep learning with different 3D convolutional neural networks (CNNs) have been used for multiclass segmentation of mandible, teeth, and background in CBCT scans. To train the models, 500 volumes were obtained from patients who had undergone orthodontic treatment in Centri Dentistici Primo company and segmentations were manually created with 3D Slicer. The segmentation performance of all trained CNNs was assessed by the DSC (Dice similarity coefficient) and RVD (Relative Volume Difference). Fully automated segmentation demonstrated a large overlap with manual segmentations (DSC: 0.9059 ± 0.0273 and 0.8939 ± 0.0325, respectively for mandible and teeth; RVD: -0.0074 ± 0.0514 and -0.0125± 0.0601, respectively for mandible and teeth). In the end, automatic segmentations of mandible and teeth have been saved in stl (Standard Triangulation Language) format and an experimental evaluation has been performed: 40 surgical templates were designed (20 patients, for each both manual and automatic segmentation). During the experiment, engineers did not know if the segmentation was manually or automatically. Subsequently, surgical guides were compared and the deviation between the two designs (manual and automatic, for each patient) was measured using the EXOCAD software, reaching a maximum deviation of (348 ± 212.78) μm. With the 3D CNN implemented, multiclass segmentation takes about 9 seconds for 1 CBCT scan, while manual segmentation takes about 40 minutes. This study demonstrates that multiclass mandible and teeth segmentation with deep learning is accurate and generally usable to reduce preoperative planning time in guided implant surgery.