机器学习辅助的通过三维面相判别骨性畸形分类的初步研究

Diagnosis test for classification of skeletal discrepancies by machine learning based on three-dimensional facial profile

目的尝试通过三种机器学习方法对三维面相进行矢状向及垂直向骨性畸形判别,对比其准确性并筛选面部软组织判别指标。方法本研究为回顾性研究。研究纳入292例正畸治疗前拍摄三维面相和头颅侧位X线片的成年患者。应用三维可变模型对三维面相数据进行面部软组织特征点标定并获得坐标,将特征点的常见几何特征和人口学特征作为待选特征,采用随机森林、自适应提升算法、多层人工神经网络对矢状向及垂直向骨性畸形进行判别,并以同一位经验丰富的正畸医师使用头影测量软件获得的结果作为金标准进行准确性比较,通过绘制受试者操作特征曲线和计算曲线下面积对各模型进行评价。计算各特征的基尼重要性,筛选出最有意义的面部软组织特征。结果对于矢状向骨型分类,多层人工神经网络的曲线下面积最大(骨性Ⅱ类0.92,骨性Ⅲ类0.97);对于垂直向骨型分类,随机森林的曲线下面积最大(低角0.84,高角0.88)。三种算法对矢状向骨性畸形的判别精度(骨性Ⅱ类:随机森林90.0%,自适应提升算法85.7%,多层人工神经网络95.0%;骨性Ⅲ类:随机森林84.6%,自适应提升算法92.3%,多层人工神经网络93.3%)均优于对垂直向骨性畸形的判别精度(低角:随机森林68.2%,自适应提升算法72.0%,多层人工神经网络76.9%;高角:随机森林76.2%,自适应提升算法77.8%,多层人工神经网络69.6%)。特征筛选结果发现"眉间点-颏前点-鼻中隔下点的角度"、"双侧太阳穴点连线中点-双侧口角点连线中点-颏前点的角度"对骨性Ⅱ类及高角的判别而言较有意义。结论机器学习方法可以在一定程度上对三维面相进行矢状向及垂直向骨性畸形的判别。

Objective This study aims to distinguish sagittal and vertical skeletal discrepancies using three machine learning methods based on three-dimensional(3D)facial profile,and a comparison among the three methods was carried out,and the discriminant features were screened.Methods Three dimensional facial profiles were collected from 292 patients before orthodontic treatment.Cephalometric analysis was conducted for diagnosis of sagittal and vertical skeletal discrepancies.The 3D morphable model method was used to obtain soft tissue landmarks.The coordinates of landmarks,along with their geometric features and demographic characteristics,were set as features for machine learning.Random Forest(RF),AdaBoost,and Multi-Layer Perceptron(MLP)were used for skeletal discrepancies discrimination.Each model was evaluated using the receiver operating characteristic curve(ROC)and the area under the curve(AUC)was calculated.Features with the highest Gini importance were selected.Results The AUC of MLP was the largest for sagittal pattern(0.92 for ClassⅡand 0.97 for ClassⅢ).For the vertical pattern,Random Forest had the largest AUC(hypodivergence 0.84,hyperdivergence 0.88).The accuracy of discrimination of the three algorithms for sagittal pattern(ClassⅡ:RF 90.0%,AdaBoost 85.7%,MLP 95.0%;ClassⅢ:RF 84.6%,AdaBoost 92.3%,MLP 93.3%)was better than that of vertical pattern(hypodivergence:RF 68.2%,AdaBoost 72.0%,MLP 76.9%;hyperdivergence:RF 76.2%,AdaBoost 77.8%,MLP 69.6%).The results of feature screening showed that the most significant features for the ClassⅡand hyperdivergence were the angle of the interbrow point to the premental point to the subnasal septum point,and the angle of the midpoint of the bilateral temple point to the midpoint of the bilateral cheilion point to the premental point.Conclusions Machine learning methods can assist in identifying sagittal and vertical skeletal deformities to a certain extent through three-dimensional images,and the features with the most significance for discrimination of skeletal patterns were revealed.