YOLO-Based Multiclass Classification of Peri-Implant Bone Loss Severity on Panoramic Radiographs Using Expert-Defined Reference Labels: A Retrospective Study

Technical Summary: YOLO-Based Multiclass Classification of Peri-Implant Bone Loss Severity

Problem Statement
The radiographic assessment of peri-implant bone loss on panoramic images is a standard clinical practice but remains susceptible to observer variability, particularly in identifying early-stage lesions. While artificial intelligence (AI) and deep learning (DL) have shown promise in dentomaxillofacial radiology, previous studies have largely focused on binary disease detection or implant identification. There is a lack of research applying object detection frameworks to the multiclass severity stratification of peri-implant bone loss on panoramic radiographs. Furthermore, many existing classification systems rely on clinical parameters (e.g., probing depth) that are unavailable in image-only datasets, necessitating a reproducible, ratio-based radiographic framework for automated analysis.

Methodology
This retrospective single-center study developed and validated a YOLO-based deep learning framework for automated multiclass classification.

• Dataset: The study utilized 1,020 panoramic radiographs containing 3,051 annotated dental implants. The data was split into training (693 images), validation (228 images), and test (99 images) sets using a stratified 70:20:10 ratio.
• Annotation & Reference Standard: Three specialists (an oral and maxillofacial radiologist, a periodontologist, and a prosthodontist) independently assessed implants based solely on radiographic findings. Disagreements were resolved by majority vote. Implants were annotated using rectangular bounding boxes on the makesense.ai platform.
• Classification Framework: Adopting a ratio-based approach derived from Hwang et al., bone loss was categorized into four classes relative to the total intraosseous implant length:
  • Healthy: No detectable bone loss.
  • Grade S: Bone loss ≤ 25%.
  • Grade M: Bone loss between 25% and 50%.
  • Grade A: Bone loss ≥ 50%.
• Model Architecture: Two object detection architectures, YOLOv8 and YOLOv12, were trained under identical conditions (300 epochs, batch size 16, Adam optimizer) using transfer learning.
• Preprocessing: The training pipeline included extensive real-time data augmentation (rotation, scaling, translation, shearing, perspective transformation, horizontal flipping) and advanced techniques like mosaic, mixup, and copy-paste augmentation to address class imbalance and improve generalization.
• Evaluation Metrics: Performance was assessed using Precision, Recall, F1-score, mAP@0.5, and mAP@0.5:0.95.

Key Results
Both models successfully detected all four severity categories, though performance varied by class and metric:

• Overall Performance:
  • YOLOv8 demonstrated higher Recall across most classes, particularly for "Healthy" (0.83) and "Grade S" (0.92), indicating strong sensitivity for dominant and early-stage categories. It also achieved the highest mAP@0.5 for the Grade A class (0.52).
  • YOLOv12 achieved higher mAP@0.5 values for the Healthy (0.46), Grade S (0.64), and Grade M (0.37) classes, suggesting a better balance of precision and localization for these categories.
• Class-Specific Findings:
  • Grade S was the most consistently detected category in both models, with the highest number of true-positive predictions in the confusion matrix.
  • Misclassification: Errors occurred most frequently between adjacent severity groups (e.g., Healthy misclassified as Grade S, or Grade M misclassified as Grade S or A).
  • Precision: Precision values were generally low across all classes for both models (ranging from 0.21 to 0.57), while recall remained higher.
• Training Dynamics: Both models showed progressive loss reduction and metric stabilization. No significant divergence between training and validation trends was observed, indicating convergence without severe overfitting.

Significance and Claims
The authors claim that the proposed YOLO-based framework demonstrates the feasibility of automated multiclass classification of peri-implant bone loss severity on panoramic radiographs.

• Clinical Utility: The strong detection performance for Grade S suggests potential value for early radiographic screening, where subtle changes are often overlooked or interpreted inconsistently by clinicians.
• Role of AI: The study positions these models as decision-support tools rather than standalone diagnostic systems. The complementary strengths of YOLOv8 (high recall) and YOLOv12 (high mAP in specific classes) support their use in enhancing workflow consistency and prioritizing cases for expert review.
• Methodological Contribution: By utilizing a ratio-based radiographic framework, the study offers a reproducible labeling strategy that mitigates the influence of panoramic magnification and distortion, making it suitable for image-only AI training without clinical variables.

Limitations Acknowledged
The authors explicitly note several constraints:

1. Imaging Modality: The use of 2D panoramic radiographs limits 3D assessment and may underestimate buccal/lingual bone loss compared to CBCT.
2. Reference Standard: Labels were based exclusively on radiographic interpretation, lacking clinical variables (probing depth, bleeding) central to definitive peri-implantitis diagnosis.
3. Generalizability: The study was single-center with a single imaging system, and no independent external validation dataset was used.
4. Class Imbalance: The dataset was imbalanced, with Grade S being dominant and Grade A/M underrepresented, which may affect model optimization for rarer classes.

In conclusion, the paper asserts that while YOLO-based models can effectively categorize peri-implant bone loss severity, their deployment requires further multicenter validation, integration with clinical data, and prospective evaluation before routine clinical implementation.