Study On The Detection And Localization Of Anatomical Markers In Cephalometric Analysis

Cephalometric analysis is often used to analyze the relationship between human and animal skulls,teeth and other bones.The location detection of marked points from cephalometric roentgenogram is regarded as an important step to study the structure of head and face,orthodontics and clinical diagnosis.With the development of deep learning technology,regions with CNN features(RCNN)and fully convolutional networks(FCN)have become powerful tools for cephalometric analysis.These tools can not only improve the detection efficiency and accuracy,but also meet the needs of modern clinical applications and provide more reliable solutions for clinical practice.In order to solve the problem of the location detection accuracy of marked points in cephalometric analysis is not ideal.First of all,a multi-task learning network is proposed on the basis of improving the signal-to-noise ratio and quality of lateral skull X-ray images by image preprocessing,including improved Faster Regions with CNN Features(Faster R-CNN)and Hourglass network.Faster R-CNN divides the target detection task into two subtasks: border regression and target classification,and Hourglass network is used to learn more accurate target position estimation.There are still missing detection at the marking points and the detection speed is slow Due to the improved Faster R-CNN model.Based on the Region-based Fully Convolutional Networks(R-FCN),a pooling layer for accurately perceiving the region of interest was adopted to effectively analyze the spatial relationship between marked points in this paper,so as to solve the problem of inaccurate target location caused by limited coordinate quantization accuracy,and adopts transfer learning technology to speed up model training and improve model performance.The success detection rate(SDR)of improved Faster R-CNN proposed in this paper for marked points reached 78.69%,83.22%,88.95% and 95.71% within the allowable error accuracy of 2.0 mm,2.5 mm,3.0 mm and 4.0 mm,respectively.The missing detection rate was 32%,and the detection time was 12.06 s.And the accuracy of 2.0 mm was improved by 0.48%,3.32% and 4.07% respectively compared with Wang,Dai and Qin’s methods.The proposed method achieved the best performance in 2.0 mm detection accuracy.In addition,the average SDR of the improved R-FCN was 84.28% under four kinds of precision,the average standard deviation(SD)was2.45 mm,the missing detection rate was 12.67%,and the detection time was 1.87 s.Compared with the improved Faster R-CNN,the missing detection rate was reduced by 19.33%,the detection of an image was shortened by 10.19 s,and the detection efficiency is improved by 84.49%.Therefore,the improved Faster R-CNN model can achieve higher detection accuracy,and the improved R-FCN model is more suitable for practical detection tasks.