Research On Liver Tumor CT Image Segmentation Method Based On Deep Learning

Liver cancer,or malignant tumor of the liver,ranks among the top three bringing about causes of death from cancer around the world,leading to a gave menace to individual health and affecting the family life of patients.Computed tomography(CT)has been extensively applied to screen,diagnose and measure tumor volume,shape and location in order to help doctors make accurate assessment and treatment at an early stage here and now.Nevertheless,traditional approaches of manually mapping liver and tumor from a great quantity of CT slices are time-consuming,laborious and highly dependent on the subjective experience of the clinician.In addition,the contrast between the liver and other adjoining viscera are lower;Liver tumors can vary in size,location,and number within a patient.Therefore,computer-aided automatic segmentation of liver and tumor in abdominal CT slices has attracted more and more attention from researchers in the field of clinical and medical imaging.With the progress of computer hardware,new methods represented by deep learning technology performed very well in computer vision,medical imaging and other fields.Among them,U-Net network is very popular in medical image segmentation,but it only uses convolutional training network in a simple stacking way,so the performance of U-Net in segmentation of complex images such as liver tumor needs to be improved.In order to improve the segmentation effect of U-Net in liver and tumor,some improvements were made to the original U-Net.First,we add residual connections at each layer of the left and right encoding-decoding branches to aid in network training.Meanwhile,unlike U-Net,which extracts feature maps from encoders and then directly copies and splicing them into decoders using skipped connections,we lead into a self-attention mechanism to capture the channel correlation between any two feature maps and update each channel map using the weighted sum of all channel maps.By including the channel attention module in the skipped connection,the redundant features learned by the corresponding encoding layer are reduced,which helps to reduce the effect of noise.In addition,we improve the bottom structure of U-Net by proposing a hybrid dilated attention convolutional block that efficaciously improves the receptive field of the model.To accomplish the goal of accurately segmenting liver tumors,in this paper,we first train the network for segmenting livers,and then use the trained liver model as a pre-trained model for tumors.The paper evaluates our advanced approach on the publicly available Li TS 2017 liver tumor segmentation dataset.Experiments show that the proposed segmentation method has achieved 0.949 and 0.799 Global Dice scores in liver and tumor segmentation tasks,respectively.Compared with other state-of-the-art techniques at present,the advanced approach has competitive performance.