Research On Multi-label Segmentation Method Of Lower Limb Bones Based On Enhanced CT Images

Computer-aided diagnosis is a very popular research direction in recent years.The study of multi-label segmentation of lower limb bones in enhanced CT images can not only serve for bone three-dimensional diagnosis and orthopedic surgery planning,but also for bone removal operations,which can be used in CT vascular analysis projects.It is used for the needs of blood vessel display and positioning,which is conducive to the discovery of early hidden vascular disease.However,the current research on bone segmentation is often for single-label segmentation of a single piece or several bones,and the research on multi-label segmentation of complete lower limb bones is still rare.Therefore,the convolutional neural network,which is popular in the field of image segmentation,is used to realize fully automated multi-label segmentation of lower limb bones in enhanced CT images.In terms of image preprocessing,in order to improve the contrast of the bones and background,a threshold cut-off process is performed.The interpolation algorithm is used for the image to unify the spacing,so that the information obtained by the same size block is roughly the same.Cropping volume of interest is used to avoid unnecessary information interference,while reducing the image size and speeding up training efficiency.The three-dimensional anisotropic diffusion filter that calculates the6-directional gradient is used to denoise the image while preserving the edge details.A three-dimensional network capable of obtaining spatial location information is selected for research,and a DSC-3DUNet network based on the 3D U-Net network is designed and implemented.In order to solve the problem that the original 3D U-Net has a shallow network structure and weak feature extraction capability,a layer of network structure is deepened to enhance the network’s ability to extract high-level semantic features.The sc SE attention mechanism module is added after the dual convolution module of each encoder and decoder to dynamically adjust the feature weights during the training process.A convolution operation is added before the jump connection,and different weights are given to the coding feature map and the decoding feature map of the same size,so that the feature fusion is more efficient.In addition,in view of the problem of fewer training samples for lower limbs,a training strategy based on the idea of transfer learning is proposed,which is to use the idea of transfer learning to pre-train the network in the head and neck data set with similar bone CT values for a period of time,and then the pre-trained model is used to initialize the network parameters of DSC-3DUNet.After that,the fusion of the results of multiple models and the post-processing operation of removing small noises on the predicted results can further improve the final segmentation effect.Finally,the performance of the method is evaluated on all 17 sets of data in the test set using the three indicators of Dice coefficient,accuracy and recall rate.The effectiveness of the designed method is proved through ablation comparison experiments and lateral comparison experiments.The average Dice coefficient of the final method reached 0.977,the average accuracy reached 0.984,and the average recall rate reached 0.972.