Multi-angle Intrahepatic Blood Vessel Segmentation Based On Multi-scale Analysis And Data Equalization

In recent years,more and more patients have liver problems,since their irregular eating and sleeping in daily life.In the liver surgical planning,it is important that the accurate and effective segmentation of blood vessels for liver computed tomography(CT)images.However,due to the complex structure of liver blood vessels,large individual differences,and low contrast of blood vessel regions,it is difficult to accurately segment liver blood vessels.With the advent of theage of big data and the development of computer technology,deep learning methods have shown its super superiority and intelligence in image segmentation filed.However,there are still many difficulties in the liver blood vessel segmentation: 1)There are many small blood vessel branches at the end of the blood vessel,and the segmentation results are poor;2)In the abdominal CT image,there are unbalanced samples in the blood vessels and non-vascular areas.In the hepatic venous system,there is the imbalance of points between the voxels in the hepatic vein and the portal vein.3)There are important context information among CT slice images due to the connectivity of blood vessels,but classical 2D methods cannot obtain these information.In order to solve above problems,the main work as follows:1)In order to improve the segmentation accuracy for small-scale blood vessels in the liver,a segmentation model combing attention mechanism,multi-scale features,deep supervision and other modules with post-processing based on the 2D encoder-decoder network structure was proposed.2)In order to solve the imbalance of positive and negative samples in abdominal CT images,a coarse-to-fine segmentation model is propoased.Coarse segmentation mainly performs blood vessel location and finds the bounding box.In fine segmentation stage,the image in the bounding box is be further segmented.In addition,in order to make the coarse segmentation more accurate,the Fixed-Point method was also used during the test.At the same time,in order to cope with the challenge of voxel imbalance between hepatic vein and portal vein,a weighted exponential log loss function is designed.3)In order to obtain context information between adjacent CT slice images,three-channel sampling and multi-angle features fusion strategies are employed.In addition,the 3D model have the ability to extract the context information,therefore,2D convolution is thrown away instead of 3D convolution.At last,a series of experiments were carried out on the hepatic vein,portal vein and venous dataset.The experimental results show that the accuracy,generalization,stability and robustness of the proposed models are better than the current state of methods,and which also show that proposed method is more efficient and superior in the liver blood vessel segmentation.