Related Experiment Video
Updated: Sep 17, 2026
Reliability of Artificial Intelligence-Based Cone Beam Computed Tomography Integration with Digital Dental Images
Published on: February 23, 2024
Automated measurement of Little's Irregularity Index on intraoral photographs using a convolutional neural network
Moritz Kanemeier1, Tuna Ergün2, Thomas Stamm2
1Department of Orthodontics, University of Münster, Albert-Schweitzer-Campus 1, Gebäude W 30, 48149, Münster, Germany. kanemeier@uni-muenster.de.
Background:
Quantification of dental irregularities is essential for assessing case severity, evaluating treatment outcomes, and aiding early detection of fixed retainer failure. Measurements are typically performed on casts or digital scans, although manual measurement on intraoral photographs has been explored to enable remote monitoring. This study evaluated automated measurement of mandibular anterior irregularity on intraoral photographs using a deep learning model.
Methods:
A dataset of intraoral occlusal photographs annotated with contact points of mandibular anterior teeth was created to train a neural network model. Individual tooth displacements and Little's Irregularity Index (LII) were computed from these contact points. Annotation reliability was assessed using intraclass correlation coefficients, and agreement between model-derived and reference LII measurements was evaluated using Bland-Altman analysis on an independent test set. A supplementary, exploratory Bland-Altman analysis was performed at the tooth level. An a priori power analysis determined the required size of the test set, based on a predefined equivalence threshold of 2 mm for LII.
Results:
The model demonstrated precise localisation of contact points, with a mean radial error of 0.55 ± 0.22 mm. Tooth displacement and LII were predicted with mean absolute errors of 0.42 ± 0.32 mm and 1.37 ± 0.89 mm, respectively. For LII, Bland-Altman analysis revealed a significant bias of 1.07 mm, with limits of agreement ranging from -1.36 mm to 3.50 mm. The exploratory tooth-level analysis yielded a bias of 0.21 mm and descriptive limits of agreement ranging from -0.74 mm to 1.17 mm.
Conclusions:
The neural network enables automated detection of mandibular anterior landmarks and estimation of LII from intraoral photographs. Since the limits of agreement were wider than the predefined equivalence threshold, the proposed model can currently only be recommended as a complementary tool.
