Deep learning-based approach for 3D bone segmentation and missing tooth region prediction for dental implant planning

Artificial intelligence has recently been used in dentistry for detection, classification and segmentation of organs and lesions.^{13,19,20,21,22}. Most published applications involve 2D radiographs, such as panoramic and periapical images, while limited studies have used 3D radiographs such as CBCT. Therefore, further AI advancements in this area would greatly benefit dentists, allowing them to use computer-assisted systems for 3D imaging procedures that require knowledge and expertise. Treatment planning is a critical step in the medical and dental fields, where an accurate diagnosis forms the basis of personalized treatment strategies. Besides the training and experience of the doctor, various other factors contribute to the treatment planning process.²³.

3D volumetric data in dentistry presents unique challenges for tooth segmentation and classification tasks. Tooth segmentation is quite difficult due to several factors. First, dental data often contains complex structures such as teeth, roots, and small bone structures, making accurate segmentation difficult due to their small size and complex shapes. Another challenge is the limited contrast and presence of noise in dental images, such as CBCT scans. This can make it difficult to distinguish between different tissues or structures, leading to lower segmentation accuracy. Dental restorations such as fillings, crowns and implants can introduce new segmentation challenges. These artificial structures can have different materials and shapes that can interfere with the segmentation process. Additionally, dental anatomy and conditions can vary significantly between individuals, which adds complexity to the segmentation task as it requires dealing with various anatomical variations.

Dental implant planning relies heavily on radiographic imaging. 3D imaging equipment should be used to examine the surgical site prior to surgery, and a series of measurements should be taken within parameters allowing for anatomical variations to help provide detailed planning.^10.24. The study evaluates vital anatomical differences that influence implant planning. Real-world cases were examined to test the effectiveness of the proposed U-Net model. Numerical analysis demonstrates that the proposed deep learning models can reduce volumetric errors to less than 1%. The bone geometry of the final implant was also satisfactory, as shown in Figure 3, which presents the model result in 3D volume views. The results of the proposed models align with manual measurements, suggesting the potential usability of the technology in implant planning.

In addition, the proposed models are computationally efficient. The training, consisting of 200 epochs, required 58 minutes, and once trained, the prediction task took only 10 seconds, making it more efficient compared to other approaches. Details of the computational parameters and training history of the proposed deep learning model are provided in the Supplementary Material.

As this study aimed to segment the missing dental bone using a conventional 2D UNET model, Moufti et al.²⁵ reported some results on the use of a 3D UNET of the MONAI framework. Compared to the results reported in Moufti et al.²⁵the dice values ​​obtained in this work systematically outperformed those obtained in Moufti et al.²⁵ Training dice consistently exceeded 90%, and test dice and total (average) dice were both above 90%. This suggests that the proposed model has better segmentation performance. Additionally, a comprehensive comparison must take into account the specific context and data set. For example, Bayrakdar et al.²⁶ proposed 3D CNN model trained on annotated CBCT images to extract features and automatically generate accurate implant plans. The accuracy rates for canal and sinus/pit detection were 72.2% and 66.4%, respectively. Park et al.²⁷ focused on developing a deep learning-based system to detect missing tooth regions in panoramic x-ray images. The authors train a deep learning model using a large dataset of annotated images. This allows the model to learn the patterns and features associated with missing teeth with an average average accuracy (mAP) of 92.14% for segmenting tooth instances and 59.09% mAP for detecting missing tooth regions. The application of transfer learning and artificial intelligence to guide implant placement in the posterior mandible has also been studied by Lui et al.²⁸ and Al-Sarem et al.^ten. He et al.²⁸ conducted an in vitro study and demonstrated the potential of pre-trained deep learning models to accurately predict optimal implant positions and help dental professionals achieve accurate and efficient implant placements. Al-Sarem et al.^ten developed a deep learning-based model to detect missing tooth positions in CBCT images, achieving high accuracy using pre-trained CNN models like DenseNet169. The model shows its potential as a time-saving tool for automated dental implant planning.

Additionally, Alsomali et al.²⁹ developed a deep learning model that can automatically identify the precise location of gutta-percha (GP) markers on cone beam computed tomography (CBCT) images. These markers are used to designate potential sites for dental implants. The researchers used 34 CBCT datasets, consisting of images of patients wearing an x-ray stent during the imaging process for implant planning. These datasets were used to train, validate and test the AI ​​model. During the training process, GP markers were manually marked on the axial images, which were then used to train the deep learning model. The effectiveness of the AI ​​model was evaluated using four CBCT datasets, and it achieved a true positive rate of 83% to accurately identify GP markers, with a false positive rate of 2 .8%. Notably, 28% of areas identified by the AI ​​model as GP markers were incorrect and 17% of actual GP markers were not detected by the model. These results highlight the limitations of relying only on axial images to train an AI model and achieve accurate performance.

Similarly, Bodhe et al.³⁰ explored the development of an AI-based system designed to assist healthcare professionals. This system determines implant types and positions, locates the mandibular canal and identifies the total number of missing teeth. These capabilities aim to improve the accuracy of dental implant procedures and maxillofacial surgeries, highlighting the potential of AI to improve patient care by assisting in decision-making in complex medical tasks.

Although these previous studies provide valuable information, it is essential to mention that many of them focus on a broader range of oral and maxillofacial structures for different purposes, such as diagnosis and generalized implant planning. However, this work focuses specifically on the segmentation of the proposed implant site areas. This approach addresses the unique challenges associated with this application, which may not be the primary focus of other previous studies. Therefore, our work has provided accurate and practical solutions that improve the accuracy and reliability of implant site planning in clinical settings.

The significance of this research lies in its revolutionary use of deep learning to improve the accuracy and efficiency of dental implant planning. The study highlights the potential of U-Net models as a useful tool in dental practices, particularly for accurately segmenting the area of ​​interest of tooth structures absent in CBCT scans. Looking ahead, this research presents new opportunities for implementing AI in dental practices. It lays the foundation for future developments that could enable real-time guidance during AI-assisted surgical procedures, as well as the integration of technologies such as augmented reality to improve implant planning. Continued improvement of these AI models is expected to significantly improve dental implant patient outcomes.

Several limitations must be addressed in this study. First, the deep learning model used to segment the missing dental bone only achieved moderate accuracy. Although the model facilitates the process, it should aim to achieve more optimal performance by accurately predicting implant outcomes. This implies that further improvements or refinements are needed to improve the accuracy of the model. Second, the reliability and accuracy of available software, such as 3D Slicer, used to segment missing dental bone may vary. Although these software tools may offer some level of segmentation, there is a risk of introducing errors or inconsistencies in the final results when using this software. Another limitation is associated with the segmentation of the region of interest. Currently, the study relies on manual cropping of volumes to isolate the region of interest for segmentation of missing teeth. This process takes time. Developing a machine learning model specifically designed to efficiently and accurately crop volumes would be beneficial, leaving only the region of interest for segmentation of missing teeth. Such a model would significantly improve the accuracy and consistency of the segmentation process.