Transfer learning predicts peri-implantitis onset on baseline intraoral radiographs

Pre-trained CNNs with non-hybrid feature extraction predicted peri-implantitis from baseline intraoral radiographs with AUC 0.74–0.83 (InceptionV3 best 0.83) across 195 implants; fine-tuning performed modestly (AUC 0.42–0.69).

• Multi-center retrospective cohort: 195 implants (146 Michigan, 49 King's College London) in 108 patients; baseline periapical radiographs were obtained within one year of prosthetic restoration.
• Patient-wise splits avoided data leakage; non-hybrid feature extraction was the most consistent transfer-learning strategy, while partial and full fine-tuning underperformed.
• Authors state larger prospective studies are needed before clinical adoption.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.