Research On CT Image Segmentation Method Based On Convolutional Neural Network

CT imaging can show the anatomical structure of internal organs and tissues.CT image segmentation aims to segment organs of interest in CT images,which can provide a diagnostic reference for physicians and improve diagnostic efficiency.With the development of deep learning,CT image segmentation methods based on deep learning can automatically extract organ features with high segmentation accuracy and fast segmentation speed.Among them,the use of convolutional neural networks(CNN)for segmentation has been the mainstream method.However,CNN based methods for CT image segmentation still face many challenges.These include over-segmentation due to the spatial distance limitation of convolutional operations,and mis-segmentation between adjacent organs when they are close together.The key to reducing CT image segmentation errors is to create long-term dependency between features and to enhance attention to critical features.In this article,the CNN-based CT image segmentation algorithm is analysed in detail.The focus is on the investigation of the network structure,and targeted solutions are proposed to mitigate missegmentation problems.The main content of the research will be as follows.1.In order to alleviate the problem of over-segmentation of target organs in CT images,a CT image segmentation model AR-UNet with double path skip connection mode is proposed.The attention and residual pathways are connected in parallel in the U-shaped structure of the model.The attentional pathway uses the attentional mechanism for building long-term dependencies between similar features by correlating features.The residual pathway captures more diverse feature information by cascading residual blocks.The problem of over-segmentation caused by blurred boundaries is reduced by using double paths.In addition,a modified bottleneck is used to fill in the lost shallow information to improve organ segmentation.In the MSDSpleen public dataset,the AR-UNet model improved the DSC metric by 1.59% and reduced the over-segmentation rate by 2.39% compared to the EBP model.The experiment shows that the algorithm can effectively reduce the oversegmentation of CT images and improve the quality of the segmentation.2.To overcome the problem of mis-segmentation between adjacent organs in CT images due to their close proximity,the Trans UNet CT image segmentation model has been improved.To build a mixed transformer layer,a mixed attention mechanism and a mixed multi-layer perceptron are incorporated into the transformer structure of the Trans UNet model.The mixed attention aims to focus on important feature information,and the mixed MLP implements the information interaction of feature sequences in the spatial and channel domains.Compared to the original Trans UNet model,the improved Trans UNet model in the Synapse public dataset shows a 2.99% improvement in the DSC index and a 12.58%reduction in the Hausdorff distance.The experimental results show that the above strategies can simultaneously reduce the mis-segmentation of multiple target organs and improve the segmentation accuracy.