Detection of mandibular molar furcation involvement on intraoral radiograph by machine learning

Abstract

Objective

Artificial intelligence (AI) is speculated to help accomplish sophisticated human tasks owing to its rapid development. Radiology is at the forefront of health care specialties that are readily accessible by AI. There are studies investigating the accuracy of AI in diagnosis of dentoalveolar pathoses. However, not much is known if AI is able to detect molar furcation involvement (FI), a condition that may have limited clinic accessibility. The objective of the study is to evaluate the detection of mandibular molar FI on intraoral radiograph by machine learning.

Study Design

The school's Axium was screened and patients with or without mandibular molar FI were enrolled. Mandibular molar periapical radiographs were cropped into single tooth images and annotated manually as healthy or FI. The images were divided as training, validation and test sets. Multiple machine learning models were trained to identify FI. Images were fed into these classifiers and their diagnostic accuracy for molar FI was evaluated.

Results

Preliminary, 55 healthy and 66 FI molars were evaluated. Twelve classifiers were modified and tested, which included neural net, RBF, SVM, Adaboost, QDA, decision tree, random forest, nearest neighbors, logistic regression, Gaussian process, linear SVM, and naive Bayes. They demonstrated accuracy in the range of 45% to 60%, with random forest performing the best with 60% accuracy.

Conclusion

Machine learning seems to be a promising tool in detection of mandibular molar FI on intraoral radiograph. Larger sample size is needed to further improve the diagnostic performance of the algorithms.