An automated diagnostic support system for jaw pathologies on panoramic radiographs: a DenseNet121-CBAM deep learning study with histopathological correlation.

OBJECTIVES

To develop and prospectively validate a novel deep learning model, integrating a DenseNet architecture with a Convolutional Block Attention Module (CBAM), for the automated classification of six common jaw pathologies from orthopantomogram (OPG) images.

METHODS

A retrospective dataset of 236 histopathologically confirmed pathological OPGs (Ameloblastoma, Dentigerous Cyst, Odontogenic Keratocyst, Radicular Cyst, Squamous Cell Carcinoma, Osteomyelitis) and 5200 normal OPGs was used for model training and pre-training. The DenseNet121-CBAM model was trained using transfer learning, data augmentation, and a weighted sampler to address class imbalance. Performance was prospectively validated on an independent cohort of 228 cases, with histopathological diagnosis serving as the ground truth.

RESULTS

On the prospective test set, the AI model achieved an overall accuracy of 92.54 % in classifying jaw pathologies against the histopathological gold standard. The model demonstrated robust class-specific performance, with F1-scores ranging from 90 % to 98 %.

CONCLUSIONS

The attention-augmented deep learning model provides a rapid and highly accurate method for diagnosing common jaw pathologies on OPGs. Its performance, validated against histopathology, establishes its potential as a reliable diagnostic support tool to enhance clinical workflows and aid in early pathology detection.