Generalized Segmentation for Maxillary Sinus and Mandibular Canal in Dental Panoramic X-rays
Published in BioImage Computing Workshop, European Conference on Computer Vision 2024 (BIC ECCV 2024), 2025
Overview
Panoramic dental X-rays are the standard first look before procedures like implant placement, but two of the structures that matter most for surgical planning — the maxillary sinus and the mandibular canal — are hard to delineate reliably. This work builds a deep-learning segmentation model that identifies both well enough to generalize across different clinical sources.
Approach
- Dataset. We constructed a labeled panoramic dental X-ray dataset for the two target structures.
- Model. A U-Net backbone with a Convolutional Block Attention Module (CBAM). The attention mechanism suppressed false positives by focusing the network on the relevant anatomy, improving precision.
- Generalization. To test robustness beyond our own data, we ran external validation on an open dataset, then introduced a modified CutMix augmentation tailored to this task. By leveraging shape bias it reduced false negatives and raised sensitivity — the augmentation and regularization strategy expanded the effective training data roughly 1200×.
Results
The model reached the best test performance on our constructed dataset while also holding up under external validation — evidence it generalizes across clinical settings rather than overfitting to one source.
Links
- Code: github.com/jaykim0413/seg-pano-extended
- DOI: 10.1007/978-3-031-91721-9_17
- Presented at the BioImage Computing Workshop, ECCV 2024.
Recommended citation: *Kim, J.W., Bae, S. (2025). Generalized Segmentation for Maxillary Sinus and Mandibular Canal in Panoramic X-Rays. In: Del Bue, A., Canton, C., Pont-Tuset, J., Tommasi, T. (eds) Computer Vision – ECCV 2024 Workshops. ECCV 2024. Lecture Notes in Computer Science, vol 15638. Springer, Cham. https://doi.org/10.1007/978-3-031-91721-9_17*
Download Paper