Revilla-León, M., Gómez-Polo, M., Vyas, S., Barmak, B. A., Galluci, G. O., Att, W., & Krishnamurthy, V. R. (2023) ‘Artificial intelligence applications in implant dentistry: A systematic review’, The Journal of prosthetic dentistry, 129(2), pp. 293–300.
INTRODUCTION
Artificial intelligence (AI) research refers to development of computer systems displaying intelligent behaviour. Machine learning is a subtype of AI where programs learn statistical trends in a data set through deliberate training (supervised or unsupervised methods), using this to recognise similar patterns in new data sets. Many studies included used a type of machine learning called “deep learning” based on artificial neural networks which teach AI to process data in similar way to the human brain.
AI grants some fascinating opportunities for development in a range of dental fields. This review analyses how AI models may be applied to implant dentistry, specifically for:
- Identifying implant types from radiographs
- Predicting osteointegration and implant success
- Optimising design
METHODOLOGY
PICO (population, intervention, comparison, outcome) methodology was used to clarify the research question (relating the goals above to AI models) and identify key terms used to search for relevant papers. 5 databases were searched for articles matching the criteria up to Feb 21st 2021 and abstracts assessed.
The text was analysed for relevance. The quality of publications was assessed using the Joanna Briggs Institute Critical Appraisal Checklist for Quasi-Experimental Studies. Two reviewers gathered data and a third resolved disagreements. This review followed PRISMA guidelines for systematic reviews and meta-analysis. Ultimately 17 studies were included (Figure 1).
Figure 1: Pie chart showing the number of studies included in the review in relation to each research outcome.
RESULTS
Seven studies developed AI models for implant recognition, most combining deep learning and CBCT to train the algorithm but using 2D radiographs as testing data. In 2003 over 2000 types of dental implants existed on the market with many more available now: the need for an accurate recognition system is clear. Comparison between studies was made difficult by variation in radiographic methods and data input e.g.. projection geometry and film speed. The studies showed a 93.8-98% accuracy in implant recognition of the 10 types assessed. Lee and Jeong (2020) compared AI recognition to dentists, finding higher specificity and sensitivity in both AI and dentist recognition when both periapical and panoramic images were used.
Seven studies looked at prediction of implant success. The AI models trained using clinical, demographic and radiographic data including variables like the presence of systemic and intra-oral conditions, local anatomy, bone levels and need for bone grafting. Many variables were not trained for including prosthesis design, implant type or genetic variables and studies didn’t use a clear definition of “implant success”. It’s clear there is a lot of difficulty in obtaining data to develop these models.
Three studies looked at optimisation of designs using AI with or without FEA (tool used to replicate and analyse how designs respond to real world conditions). Li et al. (2019) found their AI model looking at implant length, thread length and pitch showed a 36.6% decrease in stress at the implant bone interface compared to FEA design. Other studies also saw optimized implant design porosity, length, diameter, accurate prediction of the elastic modulus at the implant bone interface and improved FEA calculations. Further testing is needed to evaluate outcomes in vitro.
CONCLUSION
This review highlights how AI can’t yet safely be implemented into practice however the developments are certainly promising. The AI models recognising implant type using radiographs may help develop implant recognition software programs in future but will require substantial amounts of data.
AI models looking at osteointegration success varied from 62.4-80.5% accuracy between studies. Due to the amount of patient, operator and implant variables this will be harder to develop into a useful prognostic tool.
Studies looking at how AI can optimise implant design were however very promising, especially when compared to use of FEA alone and should be considered for further research and investment.
Research Summary Written By: Oksana Paluszkiewicz, University of Manchester – BDS3