Liver Tumor Segmentation From 3D CT Images Based On Deep Convolutional Neural Networks

The liver is the largest solid organ in abdomen of the body, which has the complicated anatomical structure with vessels system. Liver lesion is serious to threat people's health and life. In recent years, computed tomography(CT) is one of the most widely used imaging modalities for detection and diagnosis of liver tumor. The main treatments include tumor resection, intervention, ablation, etc. The size, shape, and location of tumor are required in detail before therapy in order to make a fine treatment plan. Therefore, accurate segmentation of liver tumor is an essential task for treatments and diagnosis of liver cancer.There are several major difficulties for live tumor segmentation. Firstly, the size, shape, and location of tumor is various among persons. Secondly, the boundaries between tumor and surrounding normal liver tissues are ambiguous, some tumors maybe adjacent to other organs and vessels. In addition, the great diversity of tumors' appearances and inhomogeneous density make liver tumor segmentation become a challenging task. In routine clinical practices, the tumor segmentation can be done manually by operators with good expertise and experiences, but the process is timeconsuming and has low reproducibility. A large number of researches have proposed a variety of semi-automatic and automatic segmentation methods for liver tumor. In general, semi-automatic methods require manual initialization or user interaction that leads to limitations of segmentation due to human's subjectivity and experience. Most of the automatic segmentation method based on the traditional machine learning method require a number of handcrafted features. The process of manual extraction and selection is complex and requires expensive computation. Liver tumor segmentation remains a challenging topic.In this paper, deep convolutional neural networks(CNN) was applied to segment liver tumor in CT image. Patches with different size were extracted and a variant of architectures of networks were designed to learn automatically features to obtain final segmentation. Then, Graph cut algorithm was employed to refine the CNN segmentation result. The accurate and robust segmentation for liver tumor was produced by this approach without post-processing operation. In addition, handcrafted features with automatic learning features were compared. Both of them were used to train Adaboost and random forests classifiers for liver tumor segmentation, respectively. The result showed that automatic learning features performed better than handcrafted features.The experiment results indicated that the automatic learning features obtained from the CNN model performed better than the handcrafted features. The CNN combined with graph cut model outperformed the popular machine learning algorithms and level set method, and has the potential to produce accurate and robust liver tumor segmentation.