Study On Pancreatic Segmentation Via Ringed Residual U-Net

Computer-aided diagnosis(CAD)has a wide range of clinical applications,and medical image segmentation as the first step in CAD is very important.CAD has high research value and application prospects.At present,under the medical systems of various countries,the most difficult to overcome is cancer.Among them,pancreatic cancer is the first to bear the brunt.There are many reasons.The first is that pancreatic cancer is difficult to find.As a relatively small human body Organs,it is difficult to observe pancreatic cancer through images,and pancreatic cancer has many characteristics.For example,pancreatic cancer is generally very aggressive,and cancer cells will metastasize at an early stage.Once found,they have generally become malignant cancer.In addition,the survival rate of pancreatic cancer after treatment is also very low,only 5%,so how to overcome pancreatic cancer,the biggest enemy in health,is the difficulty of current medical work.In the current medical diagnosis field,the most widely used and most mature technology is CT(Computer Tomography).CT has many advantages,such as high spatial resolution and high signal-to-noise ratio.The object of this study is also CT images,mainly including:A new pancreatic segmentation method which is called RRA-UNet is proposed.On the basis of a completely convolutional network,an attention model is added to enhance the information exchange between upsampling and downsampling.At the same time,a ringed residual module is used to solve the problem in traditional deep learning networks.The gradient disappearing problem enhances the ability to share the features of the context.In medical image segmentation,the pancreas belongs to a very small organ in the abdominal cavity.Therefore,CT has some image features,such as diverse shapes,high coincidence with the gray features of surrounding organs,and low contrast.This makes pancreas segmentation the most challenging task in medical segmentation.The best way to solve the problem of small pancreas is to add attention to the algorithm and focus the algorithm on the pancreas we studied.To solve the problem of fuzzy and complicated boundaries,the main purpose is to introduce a two-way residual model,and use deep networks and an automatic learning mechanism of reverse residuals to deal with complex features.In terms of loss function,this paper uses a new loss function to replace the Dice loss function widely used in medical images.The new loss function not only pays attention to the degree of overlap between the segmentation results and the real results,but also pays attention to the similarity of the two shapes.This article uses the NIH public pancreas CT data set(82 patient sample data),using the ten-fold cross-validation method,the average DSC coefficient is 88.32 plus or minus 2.84,which is higher than other current algorithms and has a higher robustness.Therefore,in practical applications,these methods can be used in clinical medicine to improve more reliable auxiliary diagnostic data.