Min, Sherry Yang
Zheng Zhong
Nipul Tanna
Chenshuang Li
Chun-Hsi Chung
| Min, Sherry Yang | Zheng Zhong | Nipul Tanna | Chenshuang Li | Chun-Hsi Chung |
Min Sherry Yang1,3
Zheng Zhong2, Nipul Tanna3, Chenshuang Li3, Chun-Hsi Chung3
1Periodontics, University of Pennsylvania School of Dental Medicine; 2University of California, Los Angeles, David Geffen School of Medicine; 3Orthodontics, University of Pennsylvania School of Dental Medicine
Introduction
Despite having a significant impact on the outcome and stability of orthodontic and periodontic treatment, gingival phenotype analysis has to-this-date not been integrated into any mainstream virtual treatment planning tool. The key to gingival phenotype analysis is the segmentation of the covering soft tissues from the underlying teeth and bones. However, tradition CBCT post-processing methods are inadequate to clearly mark the boundary between hard and soft tissues without human intervention. In this work, we propose a novel method for teeth and bone segmentation using a deep learning algorithm. We show that a well-trained deep learning model can accurately capture the bone and teeth geometry from CBCT images without human input. If properly integrated, this method will enable automatic gingival phenotype analysis in orthodontic treatment planning tools, which can aid clinicians to make more informed decisions during the treatment planning phase.
Methods
Seven Yorkshire pig head samples were used in this work to compare results between clinical and virtual measurements. Gingival tissue thickness was clinically measured at 3 mm and 6 mm apical to the gingival margin using an endo file and digital caliper. The clinical measurements serve as the basis for comparison against the virtual measurements obtained from the output of the deep learning model. Two deep learning models (one for teeth segmentation, another for bone segmentation) were trained in a commercially available AI software (Dragonfly), using representative CBCT images from Yorkshire pigs as inputs. After deep learning, the two models were applied to the full CBCT datasets (for all six pig head samples) to create 3D reconstructions of the teeth and bones. Intraoral scans were also performed on all pig head samples, to which the 3D bone-teeth models were manually aligned. Gingival tissue thickness is then measured virtually at locations consistent with the clinical measurements.
Results
Soft-tissue thicknesses were probed at 182 locations. The clinical and virtual measurements are strongly positively correlated (r = 0.952, p < 0.001). The average deviation between clinical and virtual measurements is 0.04 mm, 95% CI [0.01 mm, 0.08 mm]. There is no significant difference between the age of the samples and between the buccal and lingual sides of the jaws.
Conclusion
The methodology proposed in this work provides a non-invasive technique for accurately measuring the soft tissue thickness using clinical routine 3D imaging systems.