Cross-Vendor Robustness of a Hybrid Deep Learning Framework for Four-Stage Periodontitis Classification on Panoramic...

Area of Science:

• Radiology
• Artificial Intelligence
• Periodontology

Background:

• Deep learning models for periodontitis diagnosis on panoramic radiographs show promise.
• Most studies utilize single-vendor data, limiting generalizability.
• Cross-vendor generalization remains a critical, under-evaluated aspect for clinical AI deployment.

Purpose of the Study:

• To evaluate the cross-vendor robustness of a hybrid Convolutional Neural Network-Computer-Aided Detection (CNN-CAD) framework for automated four-stage periodontitis classification.
• To quantify vendor-induced domain shift in AI-based periodontitis staging.

Main Methods:

• A hybrid CNN-CAD framework was extended with a YOLO-based CNN for missing-teeth quantification, enabling four-stage periodontitis classification.
• Five hundred panoramic radiographs from three vendors were used for pooled multi-vendor training and leave-one-device-out (LODO) evaluations.
• Performance was compared using segmentation backbones (e.g., Mask R-CNN) and YOLO variants, with agreement assessed against oral radiologists.

Main Results:

• Under pooled multi-vendor training, the framework achieved high accuracy (Dice coefficients up to 0.96) and expert-level agreement (ICC=0.93, MAD=0.31).
• Leave-one-device-out evaluation revealed significant performance degradation (Dice drop to 0.75-0.86), quantifying substantial vendor-induced domain shift.
• Mask R-CNN and CNNv4-tiny demonstrated strong performance for segmentation and missing teeth detection, respectively.

Conclusions:

• The hybrid framework demonstrates expert-level performance with multi-vendor data but its robustness is challenged by vendor-specific variations.
• This study provides a benchmark for cross-vendor generalization in periodontitis staging AI and quantifies domain shift.
• Future clinical deployments necessitate vendor-aware training or domain adaptation strategies.