MorphMaskFormer: a transformer-based deep segmentation model for multi-class Demirjian stage estimation from panoramic radiographs

KIRANŞAL M., Özçelik S. T. A., Aydan T., Üzen H., Fırat H., Şengür A., ...Daha Fazla

Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, cilt.141, sa.6, ss.851-865, 2026 (SCI-Expanded, Scopus)

Özet

Objectives This study aims to develop an advanced deep learning model that automatically determines third-molar developmental stages in panoramic radiographs using the Demirjian classification, improving the accuracy and objectivity of dental age estimation for forensic and clinical applications. Study Design A total of 888 panoramic radiographs from individuals aged 7 to 30 were annotated by 2 experts based on Demirjian’s A–H staging system. The proposed model, MorphMaskFormer , is built upon the classical UNet architecture, incorporating a lightweight transformer attention module inspired by Mask2Former. The model performs both binary (tooth/background) and multi-class (A–H stages) segmentation. Its performance was evaluated using IoU, Dice coefficient, Precision, Recall, and inference time, and compared against UNet, ResUNet, DeepLabV3+, PSPNet, and SegNet. Results MorphMaskFormer outperformed all baseline models, achieving a Dice score of 0.9461, IoU of 0.8985, and the fastest inference time at 78.59 ms. In multi-class segmentation, it showed high accuracy for stages A, D, and H, with an overall component accuracy of 72.41%. Conclusions MorphMaskFormer enables precise pixel-level segmentation of dental developmental stages, reducing inter-observer variability and shortening evaluation time. Its high accuracy and efficiency make it a scalable tool that enhances diagnostic confidence and supports critical clinical and forensic age-estimation decisions.