Research On Pancreatic CT Image Segmentation Algorithm Based On Deep Learning

Diseases of pancreatic organs are a serious threat to people’s lives.At present,the segmentation and detection of lesions of pancreatic organs depend on the professional analysis of abdominal CT images by clinicians.The rapid development of new technologies and methods such as artificial intelligence and deep learning provides us with a new magic weapon to solve this problem.The way of analyzing and processing medical images using deep learning methods has attracted the attention of a large number of researchers.It is of great practical and social significance to use computer-aided segmentation methods to achieve more accurate organ segmentation accuracy than doctors’ eyes,and to provide more powerful support for early diagnosis and late treatment of pancreatic diseases.We mainly study the pancreatic CT influence segmentation algorithm and its application based on a 3Dtransformer.The research contents and innovative work are as follows:(1)We summarize the excellent and cutting-edge pancreatic organ segmentation algorithms at the present stage,introduce in detail the remarkable work of convolutional neural network and visual transformer at the stage,analyze the limitations of convolutional neural network,discuss the parts to be improved in the field of medical image processing by combining the relevant research work of convolutional neural network and cutting-edge transformer,and propose two solutions for pancreas segmentation with better performance.(2)To better combine the advantages of convolution and self-attention in transformer to overcome the challenge of pancreatic segmentation,we propose a hybrid convolutional neural network algorithm based on 3Dtransformer to extend the scope of local attention calculation in visual transformer from plane domain only to spatial domain.Combined with the convolution extraction local of small convolution kernel size,a 3D transformer is used to establish a global explicit pancreatic feature relationship,and an average pooled feature fusion module from three directions is designed in the up-sampling stage to help the network compensate for the loss of up-sampling and down-sampling information and fusion features.On the widely used National Institutes of Health pancreas dataset,the mean values of DSC,Jaccard coefficient,Precision,and Recall(± standard deviation)were 0.866±0.038,0.765±0.062,0.862±0.061 and 0.877±0.057,respectively,which verified the effectiveness of the proposed algorithm model.(3)To overcome the difficulties caused by the small size of the pancreas compared with other organs,large shape differences,and different positions in different individuals,we propose a pancreas segmentation algorithm based on a channel-level transformer in this paper.Pancreas focuses attention is used to make the network model pay biased attention to the region of interest in the pancreas during training.3Dtransformer is deployed in jump connections of U-shaped network structure,integrating all skip connections,establishing the global explicit relationship of the pancreas,and positioning more detailed position information of pancreas through a multi-level global view.A feature fusion cross-attention mechanism is designed to compensate for the loss of up-sampled and down-sampled information,which can help the network train better.For all modules of the algorithm,full four-fold cross-validation experiments were performed to verify the validity of each module,and experiments were conducted on the widely used pancreas dataset of the National Institutes of Health,the mean values of DSC,Jaccard coefficient,Precision,and Recall(±standard deviation)were 0.868±0.041,0.769±0.061,0.862±0.065 and 0.880±0.060,respectively.The results showed that Our proposed algorithm model is superior to most existing pancreatic segmentation algorithms.(4)To accelerate the implementation of medical image processing algorithms based on deep learning to clinical diagnosis,this paper combines two proposed pancreas segmentation algorithms: An end-to-end pancreas segmentation system based on a channel-level transformer and a hybrid convolutional neural network algorithm based on a 3Dtransformer.The interface display of the pancreas segmentation system and the results of various functional tests demonstrate that the system has four advantages: functionality,openness,ease of operation,and expansibility.