目的: 建立一种基于牙颌面畸形患者三维颅面特征的相似性度量模型，并通过专家相似度评分对度量模型的有效性进行检测。方法: 选取2020年1月至2022年12月在北京大学口腔医院行双颌手术及术前正畸治疗的骨性Ⅲ类牙颌面畸形患者52例，其中男性26例，女性26例，根据性别分为两组。每组各随机设置1例患者作为参考样本，该组内其余患者均为测试样本。由3位专家对测试样本与参考样本的相似度进行主观评分，评分范围为1~10分，其中1分为完全不同，10分为完全相同，设定7.5分为临床可接受的相似性结果。提取患者术前锥形束计算机断层扫描(cone beam computed tomography, CBCT) 和三维面部图像的三维硬、软组织颅面特征，包括距离、角度和三维点云特征等，采用特征选择算法和线性回归模型，并与专家相似度评分结果进行拟合，建立相似性度量模型。为验证模型的可靠性，选取14例新患者进行相似度匹配，并由专家评价匹配结果的相似度，以评价相似性度量模型的可靠性。结果: 相似性度量模型显示，面中、下颅面特征是影响颅面相似度的主要特征，包括前鼻棘点-颏下点(anterior nasal spine-menton，ANS-Me)距离、右上尖牙点至眶耳平面(right canine-Frankfurt horizontal plane，U3RH)距离、左髁顶点-左下颌角点(left superior point of condyle-left gonion, CoL-GoL)距离、左髁顶点-颏下点(left condyle-menton, CoL-Me)距离、颏前点至正中矢状面垂直(pogonion-midsagittal plane, Pog-MSP)距离、右鼻翼点-左鼻翼点(right alar base-left alar base, AlR-AlL)距离、鼻尖点-软组织颏前点-下唇点(pronasale-soft tissue pogonion-labrale inferius, Pn-Pog’-Li)交角、发际点-右侧耳屏点(trichion-right tragus, Tri-TraR)距离、左外眦点-左鼻翼点(left exocanthion-left alar base, ExL-AlL)距离、骨性面下1/3、骨性面中下2/3及软组织上唇区域等。在模型可靠性测试中，14例相似性匹配案例的平均相似度评分为(7.627±0.711)分，与7.5分差异无统计学意义。结论: 本研究使用的相似性度量模型寻找的相似案例与专家主观评价匹配度高，可用于骨性Ⅲ类患者的相似案例检索。

Objective: To establish a similarity measurement model for patients with dentofacial deformity based on 3D craniofacial features and to validate the similarity results with quantifying subjective expert scoring. Methods: In the study, 52 cases of patients with skeletal Class Ⅲ malocclusions who underwent bimaxillary surgery and preoperative orthodontic treatment at Peking University School and Hospital of Stomatology from January 2020 to December 2022, including 26 males and 26 females, were selected and divided into 2 groups by sex. One patient in each group was randomly selected as a reference sample, and the others were set as test samples. Three senior surgeons rated the similarity scores between the test samples and the reference sample. Similarity scores ranged from 1 to 10, where 1 was completely different, and 10 was exactly the same. Scores larger than 7.5 was considered as clinically similar. Preoperative cone beam computed tomography (CBCT) and 3D facial images of the patients were collected. The three-dimensional hard and soft tissue features, including distances, angles and 3D point cloud features were extracted. The similarity measurement model was then established to fit with the experts' similarity scoring by feature selection algorithm and linear regression model. To verify the reliability of the model, 14 new patients were selected and input to similarity measurement model for finding similar cases. The similarity scoring of these similar cases were rated by experts, and used to evaluate the reliability of the model. Results: The similarity metric models indicated that the features of the middle and lower craniofacial features were the main features to influence the craniofacial similarity. The main features that were related to the expert' s similarity scoring included distance of anterior nasal spine-menton (ANS-Me), distance of right upper canion point-Frankfurt horizontal plane (U3RH), distance of left superior point of the condyle-left gonion (CoL-GoL), distance of left gonion-menton (CoL-Me), distance of pogonion-midsagittal plane (Pog-MSP), distance of right alar base-left alar base (AlR-AlL), angle of pronasale-soft tissue pogonion-labrale inferius (Pn-Pog' -Li), distance of trichion-right tragus (Tri-TraR), distance of left exocanthion-left alar base (ExL-AlL), lower 1/3 of skeletal face, middle and lower 2/3 of skeletal face and upper lip region of soft tissue. Fourteen new patients were chosen to evaluate the model. The similar cases selected by the model had an average experts' similarity scoring of 7.627± 0.711, which was not significantly different with 7.5. Conclusion: The similarity measurement model established by this model could find the similar cases which highly matched experts' subjective similarity scoring. The study could be further used for similar cases retrieval in skeletal Ⅲ malocclusion patients.