Localization And Aided Diagnosis Studies Of The Skull And Bone Joints Imaging Based On Deep Learning

Medical image analysis plays an important role in medical research and clinical practice.Traditional image processing involves doctors analyzing each film,which is costly in terms of labor and time.With the increasing digitalization of imaging,automated processing of medical imaging data using artificial intelligence can assist doctors in more accurate and efficient evaluation and diagnosis.This approach can relieve the pressure on medical resources and has important research value and practical significance.In this context,this thesis mainly studies the processing method of skull and bone joint X-ray images by combining deep learning with medical image analysis,and proposes an iterative-based automatic recognition method of cephalometric landmarks and a multi-task learning-based method for identification and auxiliary diagnosis of Kashin-Beck disease.Cephalometric analysis by means of skull X-ray images is a commonly used method in modern orthodontic clinical diagnosis.The cephalometric landmarks are densely distributed and relatively large in number.In view of this characteristic,this thesis proposes a two-stage iterative detection model for cephalometric landmarks.The model first trains a deep convolutional neural network to obtain preliminary prediction areas for various landmarks.Next,for each type of landmark,a detection process based on the iterative method is used to gradually reduce the prediction range and fit the true coordinates in five iterations,which achieves more precise landmark positioning.Meanwhile,the introduction of inter-model transfer learning enables migration and sharing among models of various types of landmarks,which enhances the information exchange among models.Based on the above research,this thesis continues to propose a multi-task learning-based diagnosis model for Kashin-Beck disease.The main idea is to assist in the classification of the disease by identifying the centroids of lesion-prone areas.The models are trained simultaneously for both types of strongly related tasks via hard parameter sharing.This approach allows both types of tasks to share the useful information extracted from the hand images in the shallow network.Meanwhile,the idea of the attention mechanism is introduced.The proposed method uses the predicted lesion-prone areas to generate a heatmap that is used as a weight combined with the shallow feature map.This method further enhances the weight of the relevant characteristics of the lesion-prone areas in the diagnosis of the disease,which enables targeted processing.This thesis uses the skull X-ray image dataset proposed in the IEEE International Symposium on Biomedical Imaging and the hand X-ray image dataset of Kashin-Beck disease labeled by medical professionals to evaluate the model.The experimental results show that the model achieves 87.51%accuracy in landmark detection within the error range of two millimeters.The accuracy of detecting lesion-prone areas of Kashin-Beck disease images is 97.1%,and the accuracy of disease classification and F1 score are 92.8%and 0.871 respectively.The experimental results and comparisons show that the proposed method does improve the accuracy and stability of landmark identification and disease diagnosis.The proposed model realizes reliable cephalometric image landmark detection and a diagnostic process for Kashin-Beck disease based on hand joint images.