Liver Segmentation In CT Image Based On Semi-supervised Ladder Network

We usually need to train a traditional machine learning model with a large number of labeled samples,thereby establishing a model for predicting the mark of an unlabeled sample,namely supervised learning.Supervised learning is the most approved and widespread learning approach in machine learning.In recent years,along with the rapid development of computer technology,the collection and storage of data has become relatively easy.As a result,deep learning technology has achieved remarkable results in many areas of artificial intelligence.Most of the deep learning methods use the supervised learning mode.In practical applications,obtaining a large number of marked examples may require a lot of manpower and material resources.Semi-supervised learning is a machine learning method that uses a small number of labeled samples and a large number of unlabeled samples.The advantage of semi-supervised learning is that it reduces the reliance on manual annotation.This paper proposes a ladder network based on semi-supervised learning.This method uses unsupervised learning to supplement supervised learning through horizontal connections at each layer to achieve effective use of large amounts of unlabeled data.It was applied to the liver segmentation of abdominal CT images and achieved good results.A semi-supervised architecture based on trapezoidal networks is constructed,Multi Layer Perceptron and Convolutional Neural Network are built in,and cost parameters are used to update network parameters.The patches were extracted from the abdomen images and put into the network for training.This process will generate a large amount of redundant information.In order to reduce redundant information,the super-pixel method is used to remove information on the liver parenchyma and other organs of the abdomen,and only the liver margin region is preserved,and a rough segmentation picture of the liver contour is obtained.In order to supplement the detailed information lost in the coding channel,there are lateral connections from the noise-enhanced encoder to the decoder at each layer in the decoding channel,and each layer contributes to the cost function so that each layer of the deep network can be obtained effective learning.In order to alleviate the problem of overfitting,batch normalization is added at each level of the network to speed up network convergence.The network uses the ReLU function as an activation function to solve the problem of the gradient disappearing.The experimental data show the algorithm proposed in this paper with a high accuracy.When the labeled data accounts for 8% of the total data,the pixel classification accuracy is 91.53% and 92.04%.This paper experimentally verifies the impact of unlabeled samples on the accuracy of the algorithm,and proves the usefulness of unlabeled data in my semi-supervised ladder network.