ImplantPlanNet: A deep learning framework for automatic implant planning from preoperative CBCT images.

OBJECTIVES: Preoperative implant planning based on cone-beam computed tomography (CBCT) images supports prosthetically driven treatment but remains time-consuming and experience-dependent. This study developed and evaluated ImplantPlanNet, an automatic initial implant planning framework for single-tooth missing scenarios. METHODS: ImplantPlanNet incorporates candidate localization, local patch extraction, pose estimation from local patches, and geometric parameter recovery to estimate implant position, implant long-axis direction, length, and diameter. A dataset of 144 preoperative CBCT images from single-tooth missing sites was divided into training (n=104), internal testing (n=20), and external testing (n=20) sets. ImplantPlanNet-predicted implant plans were compared with specialist reference implant plans using three-dimensional (3D) coronal deviation, 3D apical deviation, angular deviation, dimension classification accuracy, and safety-related distance measurements. RESULTS: Internal 3D coronal and 3D apical deviations were 1.53 ± 0.80 mm and 1.77 ± 0.82 mm, respectively, with an angular deviation of 5.55 ±3.39°. Corresponding external values were 1.67 ± 1.74 mm, 2.21 ± 1.66 mm, and 6.92 ±3.41°. Length classification accuracy was 65.0% in both sets; diameter classification accuracy was 100.0% internally and 75.0% externally. Safety-related distance measurements were generally comparable between ImplantPlanNet-predicted and reference implant plans, except for a slight reduction in buccal bone plate thickness in the external testing set. CONCLUSIONS: The findings support the feasibility of using ImplantPlanNet to generate automatic initial implant plans from preoperative CBCT images for clinician review in single-tooth missing scenarios. CLINICAL SIGNIFICANCE: ImplantPlanNet may support clinician-supervised CBCT-based initial implant planning by generating proposals for single-tooth missing scenarios.