Research On Liver Tumor Image Segmentation Method Based On Deep Learning

Liver cancer is one of the most common cancers in the world,and it seriously threatens people’s lives.Liver and tumor segmentation is an important step in liver cancer treatment.Liver tumor segmentation on abdominal CT images is a routine auxiliary diagnostic method.However,there are organs with similar densities near the liver in CT images,and liver tumors have the characteristics of blurred boundaries,random locations,uncertain numbers,and different shapes and sizes.These are all unfavorable factors that restrict the effect of liver tumor segmentation.In recent years,there have been many studies on deep learning segmentation methods for CT images,but these methods have problems such as oversegmentation,under-segmentation,blurred boundaries and low segmentation accuracy in liver and tumor segmentation.In view of the above problems,this paper studies liver tumor image segmentation,and proposes a cascaded improved U-Net liver tumor image segmentation method.The main work is as follows:(1)A U-Net liver segmentation method(FRes U-Net)based on multi-scale features and residual modules is proposed.The network adopts the U-Net structure,and an improved residual module is designed to replace the ordinary convolutional layer and improve the characteristics.Utilization;redesign the skip connection so that the decoder connects the multi-scale features from the encoder,reducing the semantic difference between the encoder and the decoder;adopts the hybrid loss function of the binary cross entropy loss function and the Dice similarity coefficient loss function,alleviate the class imbalance problem and speed up the network convergence.This paper has carried out sufficient experimental verification on the Li TS dataset,and achieved 93.69%,94.87% and 87.49% on the DSC,SEN and IOU evaluation indicators,respectively,which are 5.61%,7.04% and 3.91% higher than the U-Net model baseline,respectively,and the segmentation accuracy has been significantly improved,effectively suppressing over-segmentation and under-segmentation problems.(2)A U-Net liver tumor segmentation method(DAMRU-Net)that fuses dual attention mechanism and residual module is proposed.The network integrates the dual attention mechanism in the skip connection between the encoder and the decoder,and obtains the channel and spatial correlation of features through the channel attention and spatial attention mechanism,and retains richer semantic information,enabling the network to focus on Local tumor features;use the improved residual module to replace the ordinary convolutional layer to improve the utilization of features and enhance the segmentation effect of the model.In this paper,liver tumor segmentation experiments are carried out on the basis of liver segmentation.The evaluation indicators of DSC,SEN and IOU reach 76.05%,72.29% and 67.98%,respectively,which are 6.21%,5.74% and 7.63% higher than the U-Net model baseline.the segmentation of liver tumor has a clear boundary and the segmentation accuracy is significantly improved.(3)Combining theoretical results with practical applications,a graphical interface system for liver tumor medical image segmentation is designed and implemented.The system includes three functional modules: data visualization,image preprocessing,and image segmentation.The system is simple to operate and responds quickly,and can assist doctors in observing,analyzing and quickly diagnosing liver tumor lesions in CT images,simplifying the diagnosis process and improving work efficiency.