Multimodal deep learning on panoramic radiographs and CBCT slices accurately assesses third molar-inferior alveolar canal contact

ConvNeXt-based dual-stream deep learning model combining panoramic radiographs and operator-selected CBCT slices achieved macro F1-score 0.9478 in assessing mandibular third molar-inferior alveolar canal contact (250 paired cases).

• Four widely used backbones (ConvNeXt, Swin Transformer, ResNet-50, VGG16) were systematically compared; ConvNeXt achieved the highest performance.
• SHAP-based explainability analysis indicated the model's attention focused on radiographic risk signs consistent with established anatomical landmarks.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.