Research On Liver Segmentation Algorithm Of Abdominal Image Based On Iterative Convolutional Neural Network

Liver is an important organ for human metabolism,mainly locates in the right quarter rib and upper abdomen.Liver cancer is one of the most common malignant tumors in the world,with the third highest mortality.Early diagnosis and treatment can greatly reduce the incidence rate and mortality rate of liver cancer.With the development of medical imaging technology,CT and MR have become common imaging examination methods for liver lesions.The most effective treatment for liver cancer is surgical treatment.Rapid and accurate extraction of liver tissue from medical images is the basis of successful operation.It can help surgeons to determine the liver lesion area before performing liver surgery,formulate a reasonable surgical plan for patients,reduce the risk of surgery,and preserve the healthy tissue during removing the lesion as much as possible.However,due to the irregular shape of liver,low contrast with adjacent organs and blurred edges,there are still many challenges for the segmentation of abdominal liver images.In this context,we use deep learning method to achieve automatic liver segmentation from abdominal CT and MR images,respectively.Firstly,an iterative U-Net network structure for liver segmentation of abdominal CT images is proposed in this paper.Different from the classical U-Net,this network combines the automatic context algorithm and introduces the output results of the previous network for iterative training,to compensate for the lost information in the process of downsampling.The proposed algorithm is validated on 3Dircadb and ISBI 2019 liver-chaos data sets,and the experimental results show that the proposed algorithm improves the segmentation performance on both common data sets.For the ISBI 2019 liver-chaos dataset,the Dice coefficient increased by 1.6% compared with U-Net,For the 3Dircadb dataset,the Dice coefficient increased by 1.5% compared with U-Net,The other three metrics were also better than U-Net for both datasets.Secondly,an improved algorithm based on iterative 3D UNet ++ is proposed by combining automatic context algorithm to make full use of spatial information of images.The network integrates the probability map of the previous segmentation into shallow feature map.Across the thick connection of the network,the multi-scale context information is mapped to the decoder,The proposed algorithm was applied to the ISBI 2091 liver-chaos dataset for three-dimensional liver MR image segmentation.The Dice coefficient of the proposed method reached 0.942 on the 3D liver MR dataset.Compared with 3D U-Net and 3D UNet++methods,the proposed algorithm can effectively improve the performance of liver image segmentation.In addition,we use the residual blocks as backbone to further increase the feature extraction capability of the network.The experimental results of the algorithm based on the liver MR dataset under Res Net++ architecture showed that Dice reached 0.949,which achieved better segmentation results compared with the current popular segmentation network algorithm.In summary,the convolutional neural network segmentation method based on automatic combination of context information can effectively make up for the shortcomings of the traditional U-shape framework in downsampling,and achieve good results in liver segmentation of abdominal CT and MR images.Its application to clinical engineering has certain reference value,and can provide better help for clinical diagnosis and treatment of liver diseases.