Accurate segmentation of dentomaxillofacial structures in Cone-Beam Computed Tomography (CBCT) scans poses significant technical challenges, particularly for fine anatomical details such as root apices and nerve canals. Precise delineation of these structures is essential for assessing root resorption and supporting surgical planning in digital dentistry.
This work presents a method developed in the scope of the ToothFairy3 Challenge, combining instance detection and multi-class segmentation of dentomaxillofacial structures into a single unified framework. The approach adapts a Deep Watershed technique, representing each anatomical structure as a continuous 3D energy basin that encodes voxel distances to class boundaries. This instance-aware formulation is particularly well-suited to handling the narrow, geometrically complex structures that make this segmentation task demanding.
The method was trained and evaluated on the ToothFairy3 dataset, comprising 532 CBCT scans with voxel-level annotations, achieving a mean Dice coefficient of 0.742 and HD95 of 111.13 on the test set. The implementation is publicly available.
Autors: Tomasz Szczepański, Szymon Płotka
Read the article