Objective: Obtaining detailed three-dimensional images of oral and maxillofacial structures using cone-beam computed tomography (CBCT) and analyzing them to identify the planned implant fixture position is a crucial step in contemporary implant prosth...
Objective: Obtaining detailed three-dimensional images of oral and maxillofacial structures using cone-beam computed tomography (CBCT) and analyzing them to identify the planned implant fixture position is a crucial step in contemporary implant prosthodontics. Although surgical guides can be used during implant placement, their preparation and analysis require considerable time. Therefore, the purpose of this study was to develop a deep learning model that repeatedly learns clinicians’ implant placement patterns to accurately predict fixture positions, enabling implant placement closer to the clinician’s intended target.
Methods: CBCT datasets from 66 patients who visited Seoul National University Dental Hospital for implant treatment were utilized. A 3D Cascaded Network–based preprocessing pipeline, including region-of-interest extraction, resizing, and normalization, was applied to obtain enhanced CBCT volumes. : Six implant planning landmarks defining the spatial extent of the planned implant fixture were precisely detected by selecting the region of the missing tooth and applying a landmark detection network. These landmark coordinates were converted into three-dimensional heatmaps ranging from 0 to 1, following a Gaussian distribution, and used as ground-truth labels for the 3D Cascaded Network. Performance was evaluated by calculating the mean radial error (MRE) and standard deviation between AI-predicted and ground-truth landmarks. The proposed model was compared with seven state-of-the-art architectures, including ResNet, DenseNet, EfficientNet, and SegFormer variants. Additionally, the success detection rate (SDR) under a 0.5-mm threshold was assessed across all six implant planning landmarks.
Results: The proposed model achieved the lowest mean error among all compared architectures, with an average MRE of 1.06 ± 0.45 mm across the test dataset. Notably, the mesial and buccal landmarks exhibited the smallest localization errors Under the 0.5-mm threshold, the proposed framework demonstrated higher success detection rates for several landmarks compared with CNN-based and Transformer-based models, particularly for the cervical, apical, mesial, distal, and buccal landmarks.
Conclusion: This study demonstrates that an anatomically informed, coarse-to-fine deep learning framework enables highly accurate automatic detection of implant planning landmarks on CBCT images, outperforming existing models in both accuracy and clinical applicability. The proposed method showed strong generalization across CBCT scans with diverse fields of view and anatomical variations, achieving an average error of approximately 1 mm—sufficient for practical clinical use. Automated implant position prediction has the potential to reduce the complexity of surgical guide fabrication and substantially improve clinical efficiency and consistency. Nevertheless, this study is limited to single-tooth missing cases and does not incorporate prosthetic data. Future work should include expanded datasets and multi-center validation to broaden its clinical applicability.