Deep Learning-Based Classification of Mandibular Growth-Rotation Phenotypes from Lateral Cephalometric Radiographs: A Comparative Evaluation of Convolutional Neural Networks

Background/Objectives: The cross-sectional classification of mandibular growth-rotation morphology is relevant to orthodontic diagnosis and treatment planning, but conventional structural assessment is observer-dependent. This study developed and internally evaluated deep learning models for classifying existing clockwise and counterclockwise mandibular growth-rotation phenotypes on lateral cephalometric radiographs; it did not forecast future longitudinal growth. Methods: Eight hundred untreated radiographs from individuals aged 9–20 years were labelled by consensus using Björk and Skieller morphological indicators (clockwise, n = 376; counterclockwise, n = 424). A stratified 80:20 training–validation split was used. ResNet50, InceptionV3, and VGG16 were fine-tuned after Gaussian noise reduction, min–max normalisation, and limited geometric augmentation. Results: VGG16 showed the most favourable validation accuracy–loss profile. The archived aggregate confusion matrix (n = 800) yielded an apparent accuracy of 81.13% (95% CI, 78.38–83.75) and balanced accuracy of 81.32% (95% CI, 78.62–83.96); recall was 84.57% for clockwise and 78.07% for counterclockwise phenotypes. Because the matrix included the complete labelled cohort, these values are descriptive apparent performance and not an independent test estimate. Conclusions: VGG16 can classify cross-sectional mandibular growth-rotation phenotypes with promising apparent performance. External, longitudinal, and observer-benchmark validation is required before clinical use.