Research On Automatic Segmentation Of Pancreatic CT Image Based On Deep Learning

In the field of medical research,CT is one of the most reliable diagnostic basis.Imaging technique has become an indispensable link in the early diagnosis,mid-term treatment and late rehabilitation of pancreatic cancer.However,the pancreas is located deep in the abdominal cavity,and its structure and adjacent relationship are complex.It is difficult to treat it precisely.Therefore,we should pay more attention to every step of the treatment.The most important thing is to locate and manually draw the shape of the lesion according to the different conditions of each patient,so we can work out a detailed treatment plan.However,manually locating and sketching images will not only increase the workload of doctors,but also affect the accuracy of manual segmentation.Therefore,it is very valuable and important to study the automatic segmentation of pancreatic CT images by computer.The task for segmentation of pancreatic CT image designed two deep learning models in this research.The objective of the study is to establish a convolutional neural network segmentation model to automatically identify and segment a complete pancreatic organ from CT images.The improved encoder module and attention mechanism module are written in this paper The main points of the paper are as follows:(1)The encoder-decoder structure is proved to have good performance in many fields of medical image segmentation by many experimental data.Therefore,this article improves on the U-Net model and adds an attention mechanism.Enhance image features,suppress background features,and improve the accuracy of image segmentation.Dilated Convolution is used to replace the ordinary 3 × 3 ordinary convolution to capture the spatial information of the large receptive field.The evaluation criteria in computational experiments are compared with other segmentation models.In the end,the average Dice Similarity Coefficient(DSC)of the Unet-Attention+Dilated Convolution segmentation model reached 82.54±4.83%.(2)The second segmentation algorithm is improved on the basis of the first segmentation algorithm.The optimized dual path network structure is used in the encoder to continuously explore new features and combine with the original features to make the model obtain more image details.The decoder consists of the corresponding up-sampling layer of U-net.In the original U-net,the skip connection is just a simple splicing of information,which will bring too many redundant underlying features and affect the segmentation results.The channel attention mechanism module is introduced into the feature layer connection.The attention mechanism corrects the features in the encoder by generating a weight map to effectively improve the accuracy of the segmentation target.In the NIH pancreas Dataset,the data shows that the DPN-Att-Unet + Dilated Convolution network designed in this study has the highest DSC,that is,the Dice Similarity Coefficient reached 89.81%,the lowest DSC was 72.33%,and the average DSC was 85.82 ±4.73%.The experimental accuracy exceeds the first deep learning model to the current mainstream model level.