Reliable Dental Radiograph Diagnosis via Calibrated Hybrid Representation Learning

Automated diagnosis from intraoral radiographs can enable scalable oral-health screening in settings with limited access to specialist interpretation. We present a calibrated hybrid representation-learning framework for multi-class dental radiograph diagnosis that combines a fine-tuned ConvNeXt-Tiny feature extractor with classical machine-learning classifiers trained on frozen deep embeddings. We evaluate the framework on DentIRO, a four-class single-tooth intraoral radiograph dataset, using patient-level train, validation, and locked-test splits to reduce evaluation leakage. On the locked test set, the proposed model achieves approximately 98.6% macro-F1 and 99.1% accuracy, with low calibration error and Grad-CAM evidence highlighting clinically meaningful tooth structures. These results demonstrate the potential of calibrated hybrid deep-feature pipelines for accurate, reliable, and interpretable dental radiograph analysis.