Tooth Fairy: A Cone-Beam Computed Tomography Segmentation Challenge

Many recent works in dentistry and maxillofacial imagery focused on the Inferior Alveolar Nerve (IAN) canal detection, also thanks to the recently introduced Cone Beam Computerized Tomography (CBCT) that guarantees advantages w.r.t conventional CTs (e.g. lower radiation dose, lower costs and improved spatial resolution). The three-dimensional information acquired with CBCT can be crucial to plan a wide number of surgical interventions with the aim of preserving noble anatomical structures like the inferior alveolar canal, an osseous structure of the mandible which contains the homonymous nerve, artery and vein. Identifying the canal ensures its preservation in cases of impacted third molar extraction, implant positioning or removal of cystic lesions by preventing damages to dental or neural structures that would significantly reduce the quality of life. Artificial intelligence in general, and deep learning models in particular, can support medical personnel in the surgical planning procedures by providing a voxel-level segmentation of the IAN, which is more accurate that bi-dimensional annotation commonly used in the daily clinical practice. Unfortunately, the small extent of available 3D maxillofacial datasets has strongly limited the performance of deep learning-based techniques. On the other hand, a huge amount of sparsely 2D-annotated data is produced every day in the maxillofacial practice, being the de facto standard in radiology medical centers for dentistry and maxillofacial purposes. Although the amount of sparsely labeled images is significant, the adoption of those data still raises an open problem. The incomplete detection of the nerve positioning is often sufficient to facilitate a positive outcome of surgical intervention, but it is not an accurate anatomical representation. Nevertheless, 2D annotations fail to identify a considerable amount of inner information about the IAN position and the bone structure. Additionally, deep learning approaches frame the presence of dense 3D annotations as a crucial factor, but the availability of such annotations is strongly limited by the exceptionally large amount of time required. The challenge we propose aims at pushing the development of deep learning frameworks to segment the inferior alveolar nerve by incrementally extending the amount of publicly available 3D-annotated CBCT scans. Moreover, specific task for the segmentation of bones and teeth might be addressed in further editions of the challenge