Research And Implementation Of Spinal Image Segmentation System Based On Deep Learning

Thanks to the continuous development and innovation of technology,artificial intelligence has become an integral part of our daily lives,and computers can be seen in various industries.In the medical diagnosis and treatment process,computer-aided therapy is the trend for future development.The increasing pressure and a fast-paced lifestyle have made spinal diseases a huge threat to people’s health.When computer-assisted systems deal with spinal diseases,good spine segmentation results are crucial,and accurate segmentation results can also help complete subsequent medical tasks.Therefore,researching spine segmentation methods is necessary and of great significance.With the continuous development of medicine,the demand for segmentation accuracy has exceeded the ability range of traditional image segmentation techniques.On this basis,segmentation algorithms based on deep learning have attracted people’s attention,and a spinal image segmentation system has been developed based on this,to assist users in operation and research analysis.Based on the theoretical foundation of convolutional neural networks and research on spinal image segmentation techniques,further improve the shortcomings of semantic gap and multi-scale feature extraction ability in the current medical image segmentation network U-Net,this paper integrates multi-scale feature extraction and residual connections into the U-Net network,and introduces the Swin Transformer module into the model to enhance the feature extraction and utilization capabilities of the network.As a result,an improved U-Net spine segmentation model is obtained.Experimental results show that the improved spinal segmentation model has an average Dice coefficient of about 93%,which is an improvement of approximately2% compared to the U-Net benchmark model.Compared from the perspective of model performance,the improved model’s parameter performance is also in a leading position.To better utilize the voxel information of spinal images,the improved 2D spinal segmentation model is extended to a 3D network using depthwise separable convolutions to achieve network lightweighting and solve the problem of excessive parameters in 3D networks.Experimental results show that the 3D network has higher segmentation accuracy than other comparative experimental models.The system was developed based on the requirements of the segmentation task to facilitate user operations.It includes five modules: login and registration,spinal image reading and storage,preprocessing,and spinal image segmentation,with image segmentation being the core part of the system.Users can preprocess data through this system,conveniently complete spinal image segmentation tasks,and visualize the segmentation results.The various functions of the system were tested,achieving the expected results and meeting user needs.