Objective classification of masseter muscle and deep inferior tendon morphology: A comparative performance analysis of machine learning algorithms.

OBJECTIVES

This study objectively classified masseter internal structure and deep inferior tendon (DIT) morphology using ultrasonography and machine learning on 229 images from 101 (DIT) and 97 (MUSCLE) patients.

METHODS

Following hybrid feature extraction (radiomics, LBP, HOG), patient-grouped, fold-internal SMOTE balancing, and LASSO-based feature selection, four algorithms were cross-validated under a leakage-free scheme.

RESULTS

All four classifiers achieved accuracy modestly but consistently above the one-third chance level for this three-class problem (47.9-54.3%), with no algorithm showing a consistent advantage. LASSO selection-frequency analysis identified a feature-specific asymmetry: a single GLSZM feature was selected in 46 of 50 folds for MUSCLE classification but never for DIT classification, alongside HOG-dominated features in both tasks.

CONCLUSIONS

This AI-based framework provides an objective approach for classifying masseter internal structure and DIT morphology, potentially reducing observer-dependent variability. These findings may underpin future studies investigating whether automated morphological classification can support clinical assessment and procedure planning.