Truace
Subscribe

Transfer learning CNNs predicting peri-implantitis onset from baseline intraoral radiographs

Source article: Accuracy of transfer learning in predicting peri-implantitis onset from intraoral radiograph: A multi-center longitudinal retrospective study

Abstract Background The growing application of artificial intelligence in dentistry has opened new opportunities for predicting peri-implantitis (PI) using convolutional neural networks (CNNs). Transfer learning, where knowledge from a model pre-trained on a large dataset is adapted to a new, domain-specific task, may enhance performance, particularly when medical datasets are limited. This study evaluated the ability of several pre-trained CNN architectures applied to baseline 2D intraoral radiographs to predic…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1885· cite & verify
Accuracy of transfer learning in predicting peri-implantitis onset from intraoral radiograph: A multi-center longitudinal retrospective study
"22-0008-010 Camp Lejeune" by NavyMedicine is marked with Public Domain Mark 1.0. To view the terms, visit https://creativecommons.org/publicdomain/mark/1.0/.
Trace impact readingNegative state
P 70The P score combines the specificity and measured human impact of the grounded problem claim with the strength of this Trace’s cited sources.

Both sides are scored from claims and sources, not community votes.

G 65The G score combines the specificity and measured human impact of the grounded gain claim with the strength of this Trace’s cited sources.

In brief

Researchers at the University of Michigan and King's College London tested five pre-trained convolutional neural networks on baseline periapical radiographs of 195 implants in 108 maintained patients. They compared four transfer learning strategies for predicting whether peri-implantitis would later develop.

The results suggest routine baseline images may carry predictive signal that could support more individualized maintenance and earlier prevention. Performance was only moderate and varied widely by strategy, so it remains unproven whether the approach would help in real clinical practice without prospective validation.

Main points

  1. The study was retrospective and multi-center, using data from the University of Michigan (146 implants) and King's College London (49 implants).
  2. Models were trained with patient-wise splits to avoid leakage and tested on an independent test set.
  3. Non-hybrid feature extraction gave the most consistent discrimination, and InceptionV3 reached the highest AUC at 0.83.
  4. Fine-tuning strategies generally gave AUCs of 0.42 to 0.69, which indicates weak or modest performance.

The gain

Non-hybrid feature extraction on pre-trained CNNs reached AUCs of 0.74 to 0.83 in predicting later peri-implantitis from a routine baseline radiograph, suggesting a possible aid to early risk stratification.

The problem

Fine-tuning strategies performed modestly or poorly on the same task, with AUCs generally between 0.42 and 0.69, and the authors say larger prospective studies are needed before clinical adoption.

The rundown

The five architectures were ResNet50, DenseNet, VGG16, InceptionV3 and EfficientNet. Each was combined with hybrid and non-hybrid feature extraction, partial fine-tuning of 50% of layers, and full fine-tuning.

Each implant was represented by a single periapical radiograph taken up to one year after prosthetic restoration. Performance was assessed with accuracy, AUC, sensitivity and specificity.

The authors frame the models as adjunctive tools for identifying higher-risk implants, which could inform maintenance protocols and earlier preventive interventions. They describe the performance as promising but moderate.

What this doesn’t fix

The dataset is small, with 195 implants in 108 patients, the design is retrospective, and the authors state that larger prospective studies are needed before clinical adoption.

Sources

  1. Peer-reviewedClinical Oral Investigations2026-10-19

The debate