Recognition And Classification Of Masses In Pancreatic Tumor Images Based On Deep Learning

Pancreatic cancer is a common and high mortality pancreatic tumor disease.During the diagnosis process,doctors need to search for suspicious tumor areas in the complex CT images of pancreatic tumors,which inevitably generates visual fatigue and leads to misdiagnosis.Therefore,the automatic segmentation technology of pancreatic tumor can assist doctors to quickly determine whether there are abnormalities in the patient’s pancreas and then take corresponding treatments,which has important clinical value.At present,the Trans UNet model has achieved good results in the field of medical image segmentation,but the segmentation accuracy is not satisfactory due to the small percentage of pancreatic tumors in CT images and the poor interaction between feature maps of different scales.Based on this problem,this thesis adds the measures of proposing multi-scale dense connection,cosine position encoding of polar coordinates,ROI detection module and weighted Dice loss function to the Trans UNet model in a targeted manner.In order to improve the segmentation accuracy of pancreatic tumors,the main work of this thesis is as follows:(1)To address the problem of poor interaction between different scales of feature maps of pancreatic tumors with Trans UNet model,a ST-Trans UNet model based on multiscale dense connectivity and cosine position encoding in polar coordinates is proposed.The multi-scale dense connectivity is utilized to enrich the information stickiness between feature maps and to explore the interactive association of feature map information.The cosine position encoding of polar coordinates is also used to focus on the segmented edges of pancreatic tumors and avoid redundancy of feature perception.It was demonstrated that multi-scale dense connectivity and polar coordinate cosine encoding do not conflict segmentation and the effects can be superimposed on the model enhancement,which can effectively improve the fusion effect of pancreatic tumor feature maps and capture the distance dependence between pancreatic tumor pixel points with a small proportion of pancreatic tumors in CT images.The Dice coefficient of ST-Trans UNet model on pancreatic tumor segmentation was improved by 1.7%.(2)Since pancreatic tumors account for a relatively small proportion of pancreatic CT images,it will cause pixel imbalance between pancreatic tumor feature map and background,which makes the pancreatic tumor segmentation accuracy low.To address this problem,the DRST-Trans UNet model is proposed by introducing ROI detection module and weighted Dice loss function on the basis of the existing ST-Trans UNet model.The input scale of segmented image is optimized by the effect of target detection,and the Dice loss function is improved to accurately segment the pixel edges of the target of pancreatic tumor image.The experiments demonstrate that the DRST-Trans UNet model has significantly improved the segmentation effect compared with the classical models such as Trans UNet,U-Net3+,and R-UNet,and the four improvements based on the Trans UNet model do not have conflicting segmentation on the model improvement,and the effects can be superimposed.The Dice coefficient of DRST-Trans UNet model in pancreatic tumor segmentation is as high as 89.6%,which has good segmentation effect.