Segmentation Of Critical Structures In Dental CBCT Images

Cone Beam Computer Tomography(CBCT)is one of the commonly used equipment in dental imaging diagnosis.Due to its high spatial resolution and high image accuracy,dental CBCT images have better imaging quality for the complex anatomical structures of the oral cavity and maxillofacial region which assists clinicians to observe key organizational structures such as the mandible,teeth,and mandibular canal in the maxillofacial region.The mandibular canal contains important nerves and blood vessels in the mandible.It has low contrast with the surrounding tissues in dental CBCT images and occupies a small volume,making it difficult to segment the mandibular canal.In dental CBCT images,the adjacent areas of the mandible,teeth,and mandibular canal have similar gray levels and blurred boundaries,which are the difficulties in accurately segmenting these key structures.To address the above segmentation challenges,this thesis investigates the key structure segmentation model in dental CBCT images to achieve automatic segmentation of the mandible,upper and lower teeth and mandibular canal.The main research content of this thesis is as follows:(1)Construction of a multi-label dataset of dental CBCT images.As there is no publicly available dataset of dental CBCT images,the critical structures of the collected 230 dental CBCT images are annotated in three ways in this thesis.Part of the data is manually marked for the structure of the smaller segmentation target,and part of the data is marked for the larger structure of the segmentation target combined with the image processing module in the medical image software.For the remaining images,the critical structures are automatically segmented by the artificial intelligence(AI)models and then manually corrected.The approach of combining AI models to annotate the critical structures improves the work efficiency and effectively augments the dataset.(2)Aiming at the difficulties of mandibular canal segmentation,the mandibular canal segmentation model is proposed based on Frenet frame transformed the dental CBCT images.First,the dental CBCT volume is transformed based on Frenet frame to obtain the sub-volume containing the whole mandibular canal to ensure complete 3D anatomical information.Then,to further improve the performance of the mandibular canal segmentation,clDice is used to ensure the integrity of the mandibular canal structure.Experimental results show that the transformed sub-volume in the Frenet frame can improve the Dice similarity coefficient by 0.5%to 12.1%in other advanced segmentation networks.(3)To address the difficulties of segmenting maxillofacial structures in dental CBCT images,we proposed a model to segment maxillofacial structures in dental CBCT images based on attentional mechanisms and topological constraints.The coordinate attention module is incorporated into the 3D UNet to improve the segmentation accuracy,and the topological interaction loss function is used to improve the segmentation performance of adjacent segmentation label boundaries.In addition,the geometric centroid coordinates of the teeth were predicted by heat map regression.The predicted tooth centroid coordinates and the model-predicted tooth mask were then used to segment individual teeth by marker-controlled watershed transform(MWT).The model achieved a mean Dice similarity coefficient of 98.13%for the mandible,97.30%for the upper teeth,96.94%for the lower teeth and 93.48%for the mandibular canal for the segmentation.The experimental results demonstrate that the coordinate attention module and topological interaction constraints achieve accurate segmentation of maxillofacial structures in dental CBCT images.The experimental results demonstrate that the combination of coordinate attention module and topological interaction constraints can achieve accurate segmentation of maxillofacial structures in dental CBCT images,which will further assist clinicians in the diagnosis and surgical planning of maxillofacial and oral diseases.