Research On Vertebral CT Image Segmentation Based On Attention-Driven Convolutional Neural Network

Medical image segmentation is an important foundation for the diagnosis and treatment of clinical diseases,and its automation has become an active area of scientific research.Traditional medical image segmentation methods such as Watershed algorithm and Graph Cut,which can segment human tissues to a certain extent,but the segmentation performance cannot yet meet clinical requirements,while deep learning technology can achieve more accurate results through autonomous feature learning.Therefore,this thesis takes vertebral images as the research object,and proposes a segmentation method of vertebral CT images based on attention-driven convolutional neural network.First,carry out research on the preprocessing of CT images of vertebrae.According to the characteristics of CT images of vertebrae,the vertebral information is enhanced through image enhancement and filtering operations such as window adjustment technology and Guided Filter to optimize the quality of the input image of the segmentation network,and data aug mentation is performed to avoid network overfitting.Then,carry out the research of vertebral segmentation algorithm.Based on the structure of the convolutional neural network U-Net,combined with the advanced technologies such as asymmetric convolution,pyramid pooling and attention mechanism proposed in the current deep learning,the traditional U-Net network is improved at three levels:one uses asymmetric convolution to strengthen the skeleton to improve the feature extraction ability of convolution;the other uses pyramid structure to extract multi-scale features to enhance the global understanding of the image;the third introduces channels and spatial attention to enhance the representation of useful feature information.Finally,the segmentation method is applied to the segmentation task on vertebra CT images,and the segmentation performance of the above improved strategy is verified.Through comparative experiments on public spine datasets,it is shown that the Dice coefficient and m Io U of the segmentation method proposed in the thesis have reached 82.79% and 90.72%,respectively.Compared with the U-Net-based as well as non-U-Net-based network and traditional segmentation methods,the thesis obtains better segmentation results,realize the accurate and effec tive automatic segmentation of vertebra CT images,thereby providing a new solution for the automatic development of spine image segmentation.