Research And Application Of Liver CT Images Based On Improved U-Net Model

With the improvement of people’s living standard,the number of patients with liver cancers is increasing year by year.Liver cancers have caused inestimable losses to patients and their families.Computed tomography(CT)technology is a common method of examination,which has the advantages of noninvasive,clear image and high density resolution.CT scanning technology has become a common means of examination in the diagnosis of patients with liver cancers.With the development of deep learning technology,many researchers have applied the deep learning technology in the medical image segmentation field.In the segmentation task of liver and liver tumors in CT images,the liver and liver tumors have many unique characteristics,such as variable location,different shape and overlapping boundary.In this paper,the classic U-Net network model in deep learning technology is optimized and improved,and the following research works are carried out to solve the segmentation problems mentioned above:(1)To solve the problem of liver segmentation and semantic gap in liver CT images,an embedded U-Net model is proposed for liver segmentation based on multi-scale semantic information—MSN-Net model in this paper.The model uses a novel designed embedded structure,which contains a large number of skip connections and convolution blocks.The jump connection consists of a convolution module with a larger receptive field.In addition,the multi-scale information of CT images is captured by fusing high and low dimensional feature maps among each convolution block.Finally,the weighted loss combined with Dice loss and cross entropy loss are used as the loss function of MSN-Net model.The experimental results show that the MSN-Net model not only alleviates the semantic gap problem,but also further improves the segmentation performance of the model and improves the accuracy of liver segmentation.(2)Facing the complicated task of tumor segmentation and the semantic gap caused by skip connections,a U-Net model is proposed for tumors segmentation task which based on multi-scale attention mechanism in this paper.MA-Net model can not only segment liver but also accurately segment tumor lesions.In this paper,attention mechanism is introduced into MA-Net model,and two kinds of attention modules are designed: PAB module based on location attention and MFAB module based on multi-scale channel attention.Based on the two modules of attention mechanism,MA-Net model not only alleviates the semantic gap problem,but also enhances the performance of liver tumor segmentation model.Compared with other advanced models,MA-Net model also has better segmentation performance.(3)In order to achieve the effect of combining theory with practical application,this paper designs an auxiliary diagnosis platform for liver cancer.The diagnostic platform is researched and developed based on Python and Java which are two high-level programming languages.The platform can well meet the requirements of clinicians for auxiliary diagnostic functions.At the same time,the platform has the advantages of good operation state and stability,convenient operation and beautiful appearance,and can assist doctors in clinical diagnosis and treatment.