MAGNUS-2D: A Hybrid CNN–Transformer Network with Full-Resolution Decoding for Rare-Lesion Segmentation in Panoramic Dental Radiographs
2026 International Conference on Artificial Intelligence Research and Application (ICAIRA), İstanbul, Türkiye, 2 - 04 Eylül 2026, ss.266-271, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/icaira72375.2026.11769129
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.266-271
- Karadeniz Teknik Üniversitesi Adresli: Evet
Özet
Segmenting pathological findings in dental panoramic radiographs is difficult because the clinically most valuable classes, caries and periapical lesions, are small, low in contrast and extremely rare, covering roughly one thousandth of the image pixels. Convolutional encoder–decoder networks capture local texture well but have a limited receptive field, whereas transformer models provide global context at a resolution that is too coarse for thin lesion contours. We present MAGNUS-2D, a hybrid architecture that runs a convolutional path and a vision transformer path in parallel, fuses them with cross-modal attention at the deepest bottleneck only, and decodes the fused representation back to full input resolution through learned upsampling with convolutional skip connections, so that no lesion detail is lost to bilinear interpolation. The network is trained with a class-weighted composite loss combining cross-entropy, Dice and a distance-transform boundary term. On a seven-class panoramic dataset of 4476 images, MAGNUS2D reaches a test global Dice of 0.5758, an IoU of 0.4043 and a class-balanced foreground-averaged Dice of 0.5481, outperforming eight convolutional baselines, and its lead on the class-balanced score is four times larger than on the pooled one. The gains concentrate on the two hardest classes, raising caries Dice to 0.3913 from a best baseline value of 0.3611 and periapical-lesion Dice to 0.2812 from 0.2333. An ablation shows that reducing transformer capacity and adding regularization improves every metric, which indicates that on small and heavily imbalanced datasets the useful transformer is a narrow one.