AI-based evaluation of implant abutment screw torque decay: A periapical radiograph pilot study

To develop a deep learning framework for non-invasive detection of implant abutment central screw torque decay using periapical radiographs, addressing the clinical challenge of mechanical screw loosening without invasive intervention.

A controlled ex vivo dataset (2600 radiographs of Nobel CC RP implants in porcine ribs under torque: 15-35 N·cm) and a clinical dataset (n = 501 patient radiographs) were constructed. Five transfer learning strategies across three convolutional neural network (CNN) architectures (Custom CNN, VGG16, InceptionV3) were evaluated. Optimal models were validated on an independent test set (n = 100) and benchmarked against three experienced dentists. Gradient-weighted Class Activation Mapping (Grad-CAM) visualized decision rationale.

The VGG16 model with ex vivo pre-training and clinical fine-tuning achieved 88.0% accuracy (95% confidence interval [CI]: 80.0-93.2), 88.0% sensitivity/specificity, and area under the curve (AUC) 88.0%, significantly outperforming dentists' near-random accuracy (42.0-52.0%, all p < 0.05).

(1) Transfer learning sequence critically influenced performance (ex vivo→clinical fine-tuning exceeded inverse strategy by 30.0% accuracy); (2) Dentists exhibited no diagnostic advantage over chance (F1 = 43.0%-52.0% vs. 50.0% baseline); (3) Grad-CAM confirmed model focus on biomechanically critical screw-implant interfaces.

This first AI solution enables > 88.0% accurate non-invasive detection of early torque decay, demonstrating significant potential to prevent prosthetic complications through targeted retightening.

Loosening of the tiny screws holding dental implants together is a common problem that can cause expensive repairs and hassle for patients. Currently, checking if these screws are loose requires dentists to physically tighten or loosen them, which is inconvenient and can weaken the screw over time. This study developed a new, non-invasive way to check screw tightness using standard dental x-ray pictures and artificial intelligence (AI). Researchers trained the AI system: first on detailed pictures of implant screws in a lab setting (pig ribs), and then fine-tuning it using real patient x-rays. When tested on a new set of 100 patient x-rays it had never seen before, the best AI tool (using a system called VGG16) correctly identified loose screws 88.0% of the time. This was much more accurate than experienced dentists looking at the same x-rays, who performed no better than random guessing (42.0%-52.0% accuracy). The AI reliably focused on the critical parts of the screw in the x-rays. This AI tool offers a promising way to easily monitor implant screw tightness during regular check-ups, helping dentists catch problems early before they cause bigger issues, potentially saving patients time, discomfort, and cost.