Diagnostic Accuracy of Large Language Models for Peri-Implant Marginal Bone Loss on Panoramic Radiographs

Condition(s) studied

Investigational drug(s) / intervention(s)

Multimodal large language model evaluation (ChatGPT, Gemini, Claude, DeepSeek, Grok): Each anonymized ROI image is uploaded to five multimodal large language models using an identical standardized zero-shot prompt; the model assigns a Froum and Rosen score (0-3). Outputs are compared with the clinician reference standard.

Study summary

Dental implants are a routine treatment for missing teeth, but the number of implants affected by peri-implantitis, a progressive inflammatory disease of the tissues around implants, is increasing. The main radiographic sign of peri-implantitis is marginal bone loss around the implant. Early detection and correct grading of this bone loss in proportion to implant length are important for treatment planning and prognosis.

Panoramic radiographs are widely used for general screening in dental practice, but judging bone loss on them is subjective, time-consuming and depends on the clinician's experience. Multimodal large language models (vision-language models) are increasingly able to interpret images, yet their ability to evaluate peri-implant bone loss on panoramic radiographs has not been tested.

This single-center, retrospective, observational diagnostic accuracy study will evaluate how accurately five publicly available large language models (ChatGPT, Gemini, Claude, DeepSeek and Grok) detect and grade peri-implant marginal bone loss on anonymized, cropped panoramic radiographs of functionally loaded dental implants, compared with a reference standard established by three blinded clinicians (two periodontologists and one oral and maxillofacial radiologist). Bone loss will be graded with the Froum and Rosen (2012) classification: Score 0 (no visible bone loss), Score 1 (\<25% of implant length), Score 2 (25-50%) and Score 3 (\>50%).

Approximately 784 implant sites (392 with and 392 without marginal bone loss) from existing panoramic radiographs in the hospital archive will be included. No new radiographs will be taken, no patient will be contacted and no treatment will be given; informed consent has been waived by the ethics committee. All images will be fully anonymized before evaluation.

Eligibility

Primary outcome measure(s)

• Sensitivity and specificity of each large language model for detecting peri-implant marginal bone loss — Single time point, at completion of model evaluation (approximately 8 weeks after study start); retrospective images, no follow-up
  For each of the five models (ChatGPT, Gemini, Claude, DeepSeek, Grok): sensitivity and specificity for the presence of marginal bone loss (Froum and Rosen Score 1-3 vs Score 0) on anonymized panoramic ROI images, calculated against the consensus reference standard of three blinded clinicians, with 95% confidence intervals.

Trial sites (1)

More Ankara University trials in Turkey

Other trials for the same condition