Liver Tumor Segmentation Algorithm Based On Deep Learning

Liver cancer is one of the most common malignant neoplasm and one of the most lethal malignant neoplasm in the world.It is a serious threat to people’s health.CT imaging is fast and has less artifact,so it is widely used in abdominal imaging in clinic.CT imaging also has good specificity on liver tumor and it is commonly used in the diagnosis and treatment of liver tumor.With the development of computer technology,image-guide liver tumor treatment has become a development trend.Hence,it is a significant part to segment the tumor in the image precisely in various therapeutic schedules of liver tumor treatment.Compared to manual liver tumor segmentation method,automated liver tumor segmentation method comes faster and can reduce the burden of doctors.Recently,deep learning has been commonly used in image segmentation field.However,liver tumor in CT has similar gray level to those normal liver tissue and the boundary of liver tumor is foggy.Size and shape of liver tumor is quite different in every unique case.The quantity of voxels in liver tumor as foreground is much less than that in the background,called imbalanced problem.These problems make liver tumor segmentation become a challenging task.Aim at above problems,the main innovations and contributions of this paper are as follows:1.A new cross-entropy loss function is proposed in our paper.The existing crossentropy loss function is modified by adding additive offsets and normalizing weight and feature.The new loss function gives different offsets depending on the quantity of samples in a class,which can solve the data imbalance problem.Besides,normalizing the weight and the feature in our new loss function can make the network focus on optimizing the angles between feature vectors in different classes.2.A method based on detection-segmentation network,a two-stage network,is proposed in our work.In its detection stage,the network can extract the region that includes tumor,which eliminates the interference from other tissue in the background.In the stage of segmentation,the model only segments the tumor in the candidate region,which improves the robust of the model.To improve the sensitivity of detection and accuracy of segmentation,an attention module on the feature pyramid network is added in our network,which can improve the capability of combining multi-scale features and global context information.Benefit from the improvement of the detectability,more plentiful and more accurate candidate region proposals can be provided for the segmentation network.Thus,the effect of segmentation network can be further improved.