A Bone Segmentation Algorithm For Head And Neck CT Images Based On Deep Learning

With the widespread application of medical imaging technology in clinical medicine,bone segmentation of head and neck CT images is playing an increasingly important role in medical image processing.The use of efficient head and neck CT image segmentation bone algorithms to automatically locate and segment the bone structure of the head and neck is of great significance for the clinical medical diagnosis of orthopedic surgeons,the formulation of orthopedic surgery plans,or the bone removal operations in vascular analysis.Traditional medical image processing methods are not ideal in automation and segmentation accuracy.At present,CT image bone segmentation methods based on deep learning are widely used,and the fully convolutional 3D U-Net network has shown good performance in medical image segmentation.However,the data set involved in head and neck CT image segmentation involves many types of data.It not only contains ten types of bone labels,but also requires efficient and accurate segmentation of bones and blood vessels,with complex semantic information.Using 3D U-Net network will lose part of the feature information,resulting in insufficient segmentation accuracy.A two-stage network segmentation method combining coarse segmentation and fine segmentation using 3D Res Unet network and 3D U-Net++ network is proposed.The coarse segmentation stage uses a 3D Res Unet network with residual convolution blocks to segment ten types of bones,and group normalization is used instead of batch normalization to improve the segmentation accuracy.The fine segmentation stage solves the problem of the skull and blood vessels that cannot be correctly segmented in the coarse segmentation stage.Use the skull mask obtained in the coarse segmentation stage to send enter the 3D U-Net++ network to accurately segment the skull and blood vessels.Finally,the coarse segmentation results and the fine segmentation results are combined to obtain the final head and neck CT image segmentation results.Using a variety of model evaluation methods to evaluate the two-stage segmentation model,and comparing it with 3D U-Net and other networks,shows that the two-stage segmentation model has a better segmentation effect.