Accuracy to Diagnose and Simulate Premolar Alterations Using Statistical Shape Analysis

Objectives: The present study aims to capture the key anatomical features of altered versus intact premolars by a hybrid approach using Statistical Shape Analysis (SSA) and isotopological remeshing. Additionally, the study compares performance of 4 Machine Learning (ML) algorithms in identifying or simulating tooth alterations.
Methods: A total of 113 premolar surfaces from a multicenter database were included. The surfaces underwent remeshing using the MEG-IsoQuad technique and were subsequently subjected to Principal Component Analysis. The deviation in anatomical landmark positioning was assessed by calculating the mean Euclidean distances between nodes in the original and remeshed STL files. Seven critical anatomical features were identified from each tooth's isotopological mesh, and correlations between these features and shape modes were investigated. Four ML algorithms were evaluated using a 60/40 train-test split and three-fold cross-validation to predict tooth type and alterations. A separate dataset of 20 intact teeth was used for alteration simulations, followed by an assessment of key anatomical feature differences.
Results:
The first five shape modes accounted for 85% of the total shape variability, with a landmark positioning deviation of 10.4 microns (SD=6.4). Significant correlations were found between shape modes and specific premolar morphological features. The ML models demonstrated high accuracy (>86%) and precision (>84%). In the simulation tests, discrepancies in anatomical features were less than 3%.
Conclusions: The integrated approach of isotopological remeshing and SSA proved reliable in capturing key anatomical features of premolars in a multicentric population. The ML algorithms also showed promising performance in learning and predicting both altered and intact premolar shapes, thereby offering new perspectives for future dental diagnostics and computer assisted fabrication.