基于卷积神经网络的种植体周围牙槽骨关键点识别方法研究

Convolutional neural network-based key point detection method for peri-implant alveolar bone

目的:为了准确识别种植体周围牙槽骨的关键点,提出一种基于卷积神经网络(convolutional neural network,CNN)的种植体周围牙槽骨关键点识别方法。方法:首先,收集158例成人患者种植牙术后的锥形束CT(cone beam CT,CBCT)影像学资料,选择436张种植体的冠状位、矢状位的切片图像;其次,采用高分辨力网络(high-resolution network,HRNet)进行特征提取,通过属性分解热图实现单阶段牙槽骨关键点检测,并采用局部估计精化(refinement with local estimation,LE Refinement)方法减小由于热图分辨力低引起的量化误差;最后,将LE Refinement方法与None、Upsample、Offset regression方法进行对比,以验证其对种植体周围牙槽骨关键点的识别效果。结果:LE Refinement方法识别种植体周围牙槽骨关键点的平均精度均值为85.6%,均优于None、Upsample、Offset regression方法。结论:基于CNN的种植体周围牙槽骨关键点识别方法能够较好地识别种植体周围牙槽骨关键点,可以为临床医生提供参考。

Objective To propose a convolutional neural network-based key point detection method for peri-implant alveolar bone.Methods The CBCT imaging data of 158 adult patients after implant surgery were collected,involving in the coronal and sagittal slices of 436 implants.High-resolution network(HRNet)was used for feature extraction,and attribute-disentangled heatmap was applied to achieving single-stage key point detection for alveolar bone,then the quantization error caused by low resolution of the heatmap was decreased by refinement with local estimation(LE Refinement)method.LE Refinement method was compared with None,Upsample and Offset regression methods to verify its efficacy when used for detecting the key points of peri-implant alveolar bone.Results When used for detecting the key points of peri-implant alveolar bone LE Refinement method had the mean accuracy(85.6%)higher than those of None,Upsample and Offset regression methods.Conclusion The convolutional neural network-based key point detection method for peri-implant alveolar bone behaves well and provides references for clinicians.