Automatic Segmentation Of Liver And Tumor CT Image Based On Semi-supervised Deep Learning

Liver tumor is one of the most common tumors and one of the most lethal tumors.In order to carry out accurate tumor resection,we need to accurately understand the shape,location and other information of the tumor.Therefore,not only doctors need to have clinical experience,but also the assistance of science and technology.It is necessary to extract liver and tumor region by computer-aided system.In recent years,deep learning segmentation method is widely used in medical images.Deep networks for medical image segmentation(such as liver and tumor CT images)usually require a large number of training images and their corresponding labels.However,the labeled data is usually generated by manual annotation of experts,which requires high experience of experts,so it is an extremely time-consuming and labor-consuming process.In view of the above problems,this paper studies the automatic segmentation method of liver and tumor CT images based on semi supervised deep learning.The specific research work is as follows.(1)Aiming at the problem of heavy workload of data annotation in practical applications,this paper proposes a self-training semi-supervised deep learning segmentation framework based on dense conditional random fields.This method inputs a small amount of labeled data and a large amount of unlabeled data into the network at the same time.Firstly,the segmentation model is trained based on labeled data and then predicts unlabeled data.The segmentation results are optimized with dense conditional random fields.The optimized segmentation results continue to participate in the next iteration of network training as pseudo tags.The self-training semi-supervised method combined with Dense CRF alleviates the dependence on labeled data in segmentation task.(2)In the segmentation of liver CT images,3D scSE-UNet is proposed as a segmentation network.The network combines the improved scSE-block+with 3D UNet.A parallel global max pooling layer is added to scSE-block+to enhance channel attention.The network can automatically learn effective features in a joint channel-wise and space-wise manner,which helps to preserve more edge information,and improve the segmentation accuracy.The network is trained in a self-training framework with Dense CRF.Based on the completion of liver segmentation,in order to minimize the noise outside the target region,the region outside the liver is removed,and the ROI of the liver is retained for tumor segmentation.(3)For tumor CT image segmentation,we propose CBAM-ResUNet as a segmentation network.By combining the convolutional block attention module and residual module in UNet,the network can strengthen the effective features and suppress redundant features,and ensure the integrity of feature transmission.The network is also trained in a self-training framework with Dense CRF,and a small amount of tumor labeled data is used to predict a large number of images.The experimental results show that the proposed semi-supervised learning method for liver segmentation based on 3D scSE-UNet and the semi-supervised learning method for tumor segmentation based on CBAM-ResUNet can segment CT images of liver and tumor regions very well with only a small amount of labeled data.This method can achieve comparable segmentation results with the fully-supervised segmentation method,which effectively reduces the dependence on expert labeled data in liver and tumor CT image segmentation.