Multi-structure segmentation in CBCT volumes: The ToothFairy2 challenge

Area of Science:

• Medical Imaging
• Computer Vision
• Dental Diagnostics

Background:

• Cone-beam computed tomography (CBCT) is crucial for dento-maxillofacial diagnostics.
• Comprehensive multi-structure segmentation in CBCT is time-consuming, hindering research.
• Automated segmentation methods are needed for efficient analysis of maxillofacial CBCT data.

Purpose of the Study:

• To introduce ToothFairy2, a challenge for multi-structure segmentation in maxillofacial CBCT.
• To establish a benchmark dataset and standardized evaluation protocol for segmentation tasks.
• To assess the performance of automated methods on various maxillofacial structures, including teeth and their numbering.

Main Methods:

• Development of the ToothFairy2 challenge with 530 CBCT volumes and 42 annotated classes.
• Standardized voxel-wise multi-class segmentation evaluation protocol.
• Analysis of tooth detection and FDI numbering capabilities, alongside ranking stability.

Main Results:

• High performance achieved for large structures like jawbones and pharynx.
• Challenges identified in segmenting maxillary sinuses, dental restorations, and fine structures due to class imbalance and artifacts.
• Assigning correct FDI numbers for teeth proved more difficult than delineating teeth themselves.

Conclusions:

• ToothFairy2 provides a valuable benchmark for advancing automated segmentation in maxillofacial CBCT.
• Further research is needed to address challenges in segmenting complex and artifact-prone regions.
• The released data and code will facilitate the development of robust, clinically relevant segmentation tools.