Automatic identification method for 3D mandibular landmarks based on a multi-attention stacked hourglass network
This paper addresses the current limitations in directly localizing three-dimensional(3D)mandibular landmarks using cone-beam computed tomography(CBCT)volumetric data,where there is insufficient utilization of 3D spatial information and poor recognition performance for complex key points.A new method for the automatic determination of 3D mandibular landmarks is proposed to improve the accuracy of landmark identification.This method is based on a stacked hourglass network with multi-attention mechanisms.It first generates multi-view 2D projection images from 3D CBCT data,then performs landmark predictions on these 2D images.By integrating prediction information from different views,the final 3D coordinates of the mandibular keypoints are inferred.To validate the effectiveness of the proposed method,this study used a dataset consisting of 140 real adult 3D mandibular samples,all of which were precisely annotated manually.Experimental results show that the proposed method achieves an average Euclidean distance error of only 1.16 mm in the task of 3D mandibular landmark localization,which is significantly lower than the maximum acceptable error range of 2 mm commonly recognized in medical clinical practice.The study demonstrates that the new method can effectively enhance the automatic recognition accuracy of 3D mandibular landmarks and has high practical value
万方数据
维普
