| Medical image segmentation is a key step in medical image processing and analysis.Due to the reduction of physical activity and modern sedentary office work,the spine problem has become an increasingly serious problem in modern society.Cervical spondylosis,lumbar disc herniation,idiopathic scoliosis is the common spine problems.Severe scoliosis oppresses the tissues and organs of the human body,seriously affecting the health of the human body,which often requires surgical correction,using steel nails to fix the vertebral body to maintain normal physiological curvature of the spine.Before surgical correction of the spine,it is of great significance to develop a reasonable treatment plan in advance to improve the success rate of surgery and reduce the risk of surgery.These problems such as the implant angle of the nail,the depth of implantation can be determined by measuring the information on the spine model.The higher the accuracy of the spine is,the more accurate the spinal model is obtained after the 3D reconstruction.Therefore,accurate spine segmentation is of great importance.This paper analyzes the limitations and shortcomings of traditional medical image segmentation methods.According to the features of three-dimensional CT image data,the latest medical image segmentation technology and the current hot spot of deep learning,we carry out the related research.Firstly,this paper analyzes four commonly used methods of region-based,border-based,modelbased and atlas-based segmentation,analyzing their advantages and disadvantages.Because the single-altas segmentation method is difficult to adapt to the differences of structural organization among different individuals,the paper uses a multi-atlas CT image segmentation method.First,the vertebral center and rotation of each vertebra are obtained through the preprocessing step of estimating the initial posture of the spine.Then,the segmentation results are obtained through the conventional steps of segmentation method based on multi-atlas registration,such as registration and label fusion.Finally,the similarity of the segmentation results is measured.In order to obtain better segmentation performance,the paper combines deep learning,which is the current hot spot of artificial intelligence,through drawing lessons from excellent deep learning method in the field of current medical image,such as U-Net,V-Net,3D-Unet and the latest proposed concept of dense convolution neural network(Dense Net),which have achieved very good results in image processing.We constructs a new convolutional neural network,which we call Dense V-Net,then trains the neural network using multi-atlas,and finally uses the trained model to segment spinal CT data.We compare the segmentation result with the method based on based on multi-atlas registration,as well as some ways in the literature. |
