Study On Automatic Segmentation Of Liver Tumor Based On Of MR Images

Liver cancer is one of the leading causes of cancer death in the world.The death rate and mortality are increasing year by year and are seriously jeopardizing people’s life and health.Liver resection is still the main treatment method to cure some liver cancers.Magnetic resonance imaging(MRI)is the gold standard for the diagnosis of liver cancer and it is one of the most widely used medical imaging for the treatment of liver cancer.It needs to make a perfect treatment plan before liver resection.A perfect surgical procedure requires precise information on the size,number,location of the tumors so that the liver lesion area can be accurately removed.Therefore,accurate segmentation of liver tumor is the first step for the tumor resection.However,liver tumors are difficult to be accurately segmented because of their complex shapes and positions,uneven gray scales,and blurred boundary lines with surrounding tissues.In clinical trials,it can be found that the manual segmentation method can achieve accurate segmentation result with spending much time and poor reproducibility.Many researchers have proposed some semi-automatic and automatic segmentation methods for the segmentation of liver tumors.The semi-automatic segmentation method relies on human subjectivity and prior knowledge.Automatic segmentation methods are mostly based on traditional machine learning and require manual design and extraction by experience and expensive time and computation.In the study of this paper,a U-net network model based on improved fully convolutional networks(FCN)is proposed to accurately segment liver tumors.I do some work to optimize the model.The transfer learning method is used to initialize the network’s parameters.It can accelerate the training network’s convergence and generalize the network model.A part of patients’ image data is segmented by the two semi-automatic segmentation methods based on regional growing and Snakes model.Finally,the experimental data show that the result of automatic segmentation through U-net network is more accurate than semi-automatic segmentation and its dice value reaches to 87%.The model also has more robustness.