Cephalometric Landmark Detection In X-ray Images Based On Multi-scale Feature Fusion

By identifying key points on head radiographs,physicians can perform quantitative measurements of cranial morphology to guide diagnosis and treatment planning in orthodontics,orthognathic treatment,and maxillofacial surgery,which are critical in the qualitative assessment of pathology.However,the time-consuming manual annotation and the high instability of human annotation make the automatic localization of cephalometric keypoints of great relevance.The current landmark detection in cephalometric X-ray images method is mainly based on CNN,which has some problems while achieving certain accuracy:firstly,the cephalometric X-ray images lacks rich texture information,which makes feature extraction at a single scale ineffective;secondly,CNN is difficult to obtain contextual and global information.In addition,Iandmark detection in cephalometric is a small-scale target detection problem at high resolution,using heat map regression methods can achieve higher accuracy key point localization,but there are problems of model non-differentiability and high requirements for heat map resolution.To address the above problems,the main work of this paper is as follows:1.Landmark detection in cephalometric X-ray images with fused attention mechanism and multiscale features.This approach addresses the issue of landmark detection under conditions of limited texture information by leveraging the dense connectivity structure of DenseNetl21 to initially enhance semantic information of landmarks at different scales.By introducing self-attention and cross-attention mechanisms for multi-scale feature fusion,the semantic information of landmarks is effectively extracted and filtered.2.Landmark detection in cephalometric X-ray images based on multi-scale feature learning with integral regression.Integral keypoint regression is used to address the issues of non-differentiability and high computational demand when using heatmap regression in landmark detection in cephalometric Xray images,by leveraging the advantages of both coordinate regression and heatmap regression.The method uses residual connectivity to improve the feature extraction capability of the encoder,and the obtained high-level semantic feature encoding is combined with the ECA attention module in the decoder to filter useless information and generate heat maps for integral regression by multi-scale feature fusion.The global localization model and fine localization model are obtained by training with the same network structure and different size dataset images,and the location of key points of head shadow is precisely determined by the overall and then local approach.3.System for automatic landmark detection in cephalometric X-ray images.According to the landmark detection in cephalometric X-ray images method proposed in this paper,a landmark detection in cephalometric X-ray images system based on PyQt5 and PyTorch is developed,and the training,import,testing and key point detection functions of the model are realized.The system verifies the validity and practicality of the research in this paper,and also further improves the efficiency and accuracy of medical workers.