Research On Abdominal Image Segmentation Method Based On Improved Fully Convolutional Network

There are many important organs in the abdomen,and the abdomen image is an important diagnostic basis for the diseases of abdominal organs.Therefore,in clinical practice,the automatic segmentation algorithm of organ and tumor has put forward high requirements However,there are many difficulties and challenges in the research of medical image segmentation:serious imbalance and small difference between classes in medical images;insufficient annotation data;diversity of the number,shape,size and location of tumors,which easily lead to missed detection and poor segmentation effectWith the development of artificial intelligence and computer vision technology,the computer-aided diagnosis technology based on medical image has entered the stage of rapid development.Machine learning,especially deep learning,plays an important role in medical image recognition,classification and registration.With the advent of fully convolutional network in semantic segmentation task,models based on FCN are proposed for medical image segmentation,and U-Net is one of the most successful models.However,U-Net and variations of U-Net always sacrifice feature resolution to pursue high-level features.Additionally,capacity of medical image segmentation models to capture multi-scale features and generate long-range dependencies are rarely concernedBecause of the above problems,there is no assistant diagnostic tool which can be effectively used in clinical organs and tumor segmentation.This paper studies abdomen image segmentation method based on fully convolution neural network.The main work includes(1)Research on abdominal image segmentation based on context splicing network.In medical image segmentation,it is critical to extract multi-scale features and context information Based on the research of existing semantic segmentation methods,this paper proposes a context splicing network(CS-Net)to extract rich context information while retaining spatial information;(2)Construction a contour aware network(CA-Net).In order to further improve the recognition ability of the model to the liver and tumor edge,the edge extraction branch is added to the segmentation model(3)Data collection and experimental verification.Collect and process the data set provided by the hospital,standardize the label value marked by doctors,and currently 438 scanning sequences are used in the experiment.The proposed methods are evaluated on three public datasets(CHAOS,LiTS and KiTS)and hospital datasets,and compared with the existing methods.The experimental results show that the proposed models outperform U-Net and the recent proposed CE-Net and Deeplab v3+ for liver and kidney segmentation,and has competitive tumor segmentation results.