Liver Segmentation From CT Images Based On U-Net

Among the most common malignant tumors in the world,primary liver cancer ranks fifth,and the mortality rate ranks third.Liver transplantation and surgical resection are the most effective radical treatments for primary liver cancer.Before the operation,the doctor needs to obtain the anatomical information and quantitative information of the liver through medical imaging technology for liver function assessment.Usually it requires a doctor with relevant expertise and extensive practical experience to manually segment the liver region from the medical image frame by frame and pixel by pixel.However,this process is time consuming and laborious.And affected by the subjective factors of the doctor,the liver region marked by different doctors will also have a certain degree of difference.Therefore,in order to reduce the workload of doctors and speed up the efficiency of segmentation it is necessary to use artificial intelligence technology to segment liver automatically.Following discussion on the development history and existing challenges of medical image segmentation methods,this dissertation focuses on multi-task learning and weak supervised learning.The main results and contributions of this dissertation are as follows:(1)A liver segmentation network model with classification module(USCNet)is proposed.The model adopts a classification module based on U-Net to achieve multi-task learning.When segmenting the CT(Computed Tomography)image,it will judge whether there is a liver region in the image,and then post-process the segmentation result according to the classification result.Through multi-task learning,the model can learn better liver characteristics.Experiments show that the multi-task segmentation model proposed in this dissertation can better segment the liver and has better stability when dealing with different individuals.(2)Since pixel-level labels are rare,we propose a weakly supervised segmentation method(WUSCNet)which enhances segmentation through the data with image-level labels.The model is first trained by data with pixel-level labels.And after a number of iterations,additional data with image-level labels is added for training.The method extracts the regional confidence map from the data with image-level labels by CAM(Class Activation Mapping),and then uses the regional confidence map to constrain the segmentation error.The experimental results show that the segmentation performance of the model can be improved by this method because the information of the data with image-level labels is introduced.