# Detection and Classification of Peri‐Implant Marginal Bone Loss in Cone‐Beam Computed Tomography Using a Deep Learning Approach

## TL;DR

A deep learning model using YOLOv8 can accurately detect and classify bone loss around dental implants in 3D cone-beam CT images, with high performance for mild cases but lower accuracy for severe cases.

## Contribution

This study introduces a YOLOv8-based deep learning model for automated detection and grading of peri-implant marginal bone loss in cone-beam CT images.

## Key findings

- The YOLOv8 model achieved an overall accuracy of 0.90 and F1-score of 0.90 for mild bone loss cases.
- Performance dropped for moderate and severe cases with F1-scores of 0.62 and 0.70, respectively.
- The model demonstrated excellent reliability with a Kappa score of 0.954 in a three-class grading scheme.

## Abstract

Modern dental implants have high long‐term survival, but peri‐implant marginal bone loss remains a multifactorial cause of implant failure and often is radiographically occult. Cone‐beam computed tomography (CBCT) provides superior 3D assessment but produces large datasets requiring expert interpretation. Deep‐learning object‐detection models like YOLOv8 may automate detection and grading. This study aimed to evaluate a YOLOv8‐based model for automated detection and grading of peri‐implant marginal bone loss on 2D images derived from CBCT.

This retrospective study used 699 2D CBCT sections. Marginal bone loss was graded into four classes (≤ 20%, 21%–40%, 41%–60%, > 61% of implant length). A YOLOv8 detector was trained on 600 × 600‐pixel images (bounding box width 20 px) with two classes (implant, bone loss), split 80/10/10 (train/test/val), 200 epochs, batch size 8 and tuned hyperparameters. Performance was assessed by accuracy, precision, recall, and F1‐score.

The YOLOv8 model achieved strong diagnostic performance on the test set, with overall accuracy, precision, recall, and F1‐score of 0.90. It performed best for healthy sites (precision 0.92, recall 0.99, F1 0.95) and maintained high performance for mild lesions (F1 0.90), while moderate and severe cases showed reduced metrics (F1 0.62 and 0.70, respectively). A three‐class scheme had 0.88 accuracy and excellent reliability (Kappa = 0.954). Detection metric included mAP@0.5 (mean precision at intersection over union threshold of 0.5) of 0.889, a recall of 0.98 at low threshold, and implant‐length detection accuracy of 1.00. Training stabilized after about 15 epochs.

The YOLOv8‐based deep learning model can reliably detect and grade peri‐implant marginal bone loss on CBCT images. Future research should expand datasets, incorporate multimodal information, and validate performance across diverse clinical settings.