Research On Method Of 3D Segmentation For Spinal Canal CT Images Based On Sparse Annotation And Deep Learning

The incidence of spine-related diseases is getting higher and higher,which has a great impact on people’s work and life.At present,the diagnosis of lumbar spinal stenosis requires doctors to manually separate the spinal canal region for diagnosis,which is time-consuming and labor-intensive and is greatly influenced by subjective factors.Therefore,it is very meaningful to use computer technology to assist diagnosis and treatment,however,there are not all bone structures around the spinal canal,many soft tissue structures inside and around the spinal canal and the grayscale difference between the inner and outer is not obvious,it is not suitable to use the threshold method for segmentation.The contour model method often occurs over-segmentation,and the design of feature extraction operators based on image texture features is very time-consuming and labor-intensive.The contour model method often results in over-segmentation,and the design of feature extraction operators based on image texture features is very time-consuming and labor-intensive.In order to overcome the limitations of these traditional methods,this study uses a deep learning-based spinal CT image segmentation method.The 3D Attention U-net model is improved to achieve better performance on spinal canal segmentation,and two methods: fully-automated segmentation and semi-automated segmentation are proposed,which make it possible to obtain good segmentation performance for different spinal canal data samples.The specific research work is as follows:(1)First,improve the network structure of 3D Attention U-net model by adding a designed projection attention module before the attention module.When the skip connection is performed,the feature map from the coding area enters the projection attention module first to strengthen its own feature expression,and then enters the AG module to further suppress the activation of irrelevant signals and highlight the area of interest,which retains the low-level surface features and high-level abstract features of interest effectively,so that improves the segmentation accuracy of the model successfully.Then improve the loss function by introducing the parameter λ to give the cross entropy Loss and Dice Lose different weights,which can further improve the spinal canal segmentation ability of the model.(2)According to the characteristics of 3D spinal canal CT images,the sparse annotation method is improved,then two 3D neural network model segmentation methods based on sparse annotation datasets are proposed: fully-automated segmentation method and semiautomated segmentation method.The fully-automated segmentation method can predict other data samples in batches after training the model,which is very efficient.However,due to the large individual differences in the degree of curvature of the spinal canal and the shape at different positions,it is not applicable to some data samples,but the semi-automated segmentation method can successfully solve this problem and further improve the segmentation accuracy.Then,the results obtained by the fully-automated segmentation method and the semi-automated segmentation method are compared with the segmentation results of the 2D neural network models,from which we can know that the semi-automated segmentation method benefits from the similarity between layers of 3D medical image data,not only avoids the unstable results of fully-automated segmentation method,but also can achieve higher-precision prediction of the unlabeled parts of itself for different samples.