AI Prediction of Bone Graft Integration Success Using CBCT Datasets in Simulated Peri-Implant Defect

The integration of bone grafts is a determinant factor for achieving the success of implants and the outcome of healing is hard to predict. Artificial intelligence (AI) holds potential in the cone-beam computed tomography (CBCT) image analysis to predict regenerative outcomes. The purpose of the present study was to create and test a deep learning model to engage the success of bone graft integration in the case of simulated peri-implant defects in CBCT datasets.

We conducted a retrospective study based on data of 847 CBCT scans of patients who experienced loss of soft tissues around the implant due to bone grafting operations. The 12-month radiographic and clinical outcomes have been used to classify defects into successful (n=512) and unsuccessful integration (n=335). ResNet-50 built on transfer learning was proposed as a convolutional neural network (CNN). The data were separated into training (70%), validation (15%) and testing (15%) sets. The performance of the models was compared in terms of accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC-ROC) and F1-score.

The AI model had an overall accuracy of 87.3 ± 2.1, a sensitivity of 89.6 ± 1.8, a specificity of 84.2 ± 2.4 and an AUC-ROC of 0.912 ± 0.023. It was found that defect morphology classification had a high predictive ability and circumferential defects exhibited a better prediction ability (91.2) than dehiscence-type defects (82.7) (p <0.01). The model showed better results than clinical assessment (87.3% vs. 71.4% p<0.001).

The created AI model exhibits strong predictive power of bone graft integration success based on CBCT scans, which could provide a means to plan the treatment individually and make more effective clinical decisions in the field of implant dentistry.