High-frequency enhanced vision transformer for automated sagittal skeletal malocclusion classification and orthodontic screening prioritization on lateral cephalograms

Abstract

Accurate identification of sagittal skeletal malocclusion on lateral cephalograms (LCRs) is essential for orthodontic screening and timely referral. However, conventional deep learning models often struggle to capture high-frequency anatomical structures and show limited cross-dataset performance across heterogeneous datasets. Standard Vision Transformers (ViTs) may also underemphasize fine cortical bone boundaries critical for cephalometric analysis. This study proposes a High-Frequency Enhanced Vision Transformer (HFE-ViT) to improve image-based sagittal skeletal classification, probabilistic calibration, and screening-prioritization interpretation. Two public datasets: Dicle ( n = 856, ANB-based labels) and ISBI ( n = 400, landmark-derived labels) were analyzed following standardized preprocessing. HFE-ViT integrates a frequency-domain enhancement branch within each Transformer encoder using FFT-based filtering to emphasize cortical bone and dentition boundaries. Performance was evaluated using stratified five-fold cross-validation, bidirectional zero-fine-tuning external validation, label-heterogeneity sensitivity analysis, and a leakage-controlled combined-dataset training experiment in which each reported test subset was locked before any model training, validation, model selection, or calibration. Calibration used temperature scaling fitted only on training-derived validation/calibration subsets. A continuous referral-likelihood score, defined as 1 - P(Class I), was evaluated using ROC and exploratory decision curve analysis; this score represents the model-estimated probability of a non-Class I sagittal skeletal pattern and serves as a classifier-derived screening-prioritization index. A total of 1,256 subjects were included. HFE-ViT achieved internal-validation AUCs of 0.897 and 0.879 for the Dicle and ISBI datasets, respectively. Leakage-controlled combined-dataset training yielded AUCs of 0.879 and 0.885 on the two locked target test subsets. Focal loss modestly increased Class II recall in the ISBI dataset. Calibration improved after temperature scaling (ECE: 0.048 → 0.029). Grad-CAM demonstrated 76–82% alignment with expert-defined anatomical regions. The referral-likelihood score achieved AUCs of 0.902 and 0.887 in internal and external validation, with decision curve analysis suggesting exploratory potential screening net benefit for the non-Class I classification endpoint. The proposed HFE-ViT improves image-based sagittal skeletal malocclusion classification on LCRs by combining global attention modeling with high-frequency boundary enhancement and provides supportive interpretability evidence consistent with clinically relevant anatomical regions. The calibrated referral-likelihood score may support orthodontic screening prioritization, although prospective multicenter validation remains necessary before clinical implementation. Workflow for lateral cephalogram-based sagittal skeletal malocclusion classification, external validation, calibration, and referral-likelihood score evaluation using a high-frequency enhanced Vision Transformer.

The authors' abstract, as published at the source. BMC Oral Health, 2026 · DOI ↗