Multi-output classification of dental implant placement parameters in the posterior maxilla from CBCT images using a two-stage vision transformer framework

Abstract

Implant placement in the posterior edentulous maxilla is clinically challenging due to anatomical variability, limited residual bone height, and maxillary sinus pneumatization. Conventional planning relies heavily on clinician experience and manual radiographic interpretation, leading to variability and subjectivity. This study developed and evaluated a two-stage Vision Transformer (ViT)–based deep learning framework for multi-output classification of implant placement parameters from cone-beam computed tomography (CBCT) images. A retrospective dataset of 457 expert-validated posterior maxillary edentulous cases was collected from the Faculty of Dentistry, Mahidol University. The model simultaneously classified four planning parameters—implant height, implant diameter, sinus lift technique, and sinus lift stage—using a multi-output architecture with task-specific classification heads. A two-stage fine-tuning strategy was employed to optimize transfer learning. The proposed ViT model achieved an overall accuracy of 85.56%, with task-specific accuracies of 88.89% for implant height, 80.00% for implant diameter, 84.44% for sinus lift technique, and 88.89% for sinus lift stage, outperforming the best CNN baseline (62.22%) by 23.34 percentage points. The model demonstrated near-real-time inference (0.037 s per sample). These findings suggest that ViT-based multi-output classification with two-stage fine-tuning can serve as an effective decision-support tool for dental implant planning.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗