AI-driven prediction of dental implant numbers to be placed for patient-specific treatment planning.

Introduction and aims: The growing prevalence of edentulism, particularly in aging populations, has increased the demand for accurate and consistent dental implant treatment planning. Determining the number of implants required in edentulous areas is a complex process influenced by various clinical and anatomical factors. Traditional approaches rely heavily on clinician experience and image interpretation, often resulting in variability. This study aimed to develop and evaluate a deep learning-based regression model, using panoramic radiographs and clinical data, to predict the optimal number of implants required for edentulous patients and to support standardised, data-driven treatment planning.

Methods: This retrospective study included 628 patients (341 females, 287 males; mean age 60.0 ± 14.7 years) treated with dental implants at Daejeon Dental Hospital, Wonkwang University between 2019 and 2023. A total of 919 edentulous regions of interest (ROIs) were labeled with the number of implants determined by consensus between 2 oral and maxillofacial surgeons using panoramic and CBCT imaging. Preprocessing involved ROI extraction, zero-padding, resizing, normalisation, and channel duplication. A Vision Transformer (ViT)-based regression model was constructed and trained using transfer learning from the pre-trained ViT-Base (google/vit-base-patch16-224-in21k) model. The [CLS] token output was passed through a custom regression head to predict implant counts. Model training employed 5-fold cross-validation with the Adam optimiser, Mean Squared Error (MSE) loss function, and dropout for regularisation. Performance was evaluated using MSE, mean absolute error (MAE), R², and explained variance score (EVS), complemented by visual diagnostic plots.

Results: The ViT-based model achieved strong predictive performance with an MSE of 0.0460, MAE of 0.0871, and both R² and EVS values of 0.9189. Visual diagnostics (residual, box, and Q-Q plots) confirmed that the model's errors were symmetrically distributed and approximately normal, indicating good model fit and reliability. The model effectively captured spatial relationships within the edentulous areas and demonstrated potential as a quantitative decision-support tool for implant planning.

Conclusion: This study demonstrated the feasibility of using a deep learning-based ViT regression model to predict the number of dental implants needed in edentulous patients. The model showed high accuracy and explanatory power, suggesting its potential utility in standardising implant treatment planning. However, the study's generalisability is limited by its single-center design and sample size. Future work should incorporate multi-institutional data, 3D segmentation, and additional clinical variables to further improve model accuracy and clinical applicability.

Clinical relevance: The proposed AI model provides a reliable, data-driven approach to assist clinicians-especially those with less experience-in determining implant quantity for edentulous patients. Our model the potential to enhance consistency in treatment planning, reduce variability and ultimately improve patient care in implant dentistry.