Research On Automatic Segmentation Method For Liver CT Images Based On Semi-supervised Learning

Medical image segmentation is one of the most important steps in computer-aided intervention and diagnosis,but it is still difficult to train high-accuracy deep neural networks in the field of medical image processing.First,manual labeling of medical pixel-level labels is time-consuming and tedious,and requires domain knowledge,so the amount of labeled medical image data is generally small.Therefore,effective use of unlabeled data is crucial for medical image segmentation.Second,in the Computed Tomography images of the abdomen,due to the low contrast between the liver and adjacent organs,changes in the shape of the liver,the difference in the contrast between the liver and the internal tumor tissues,and the noise in the CT images,the segmentation of the liver CT images has become a challenging task.In view of the above issues,this paper proposes two automatic segmentation methods for liver CT images based on semi-supervised learning: semi-supervised liver CT image segmentation method based on deep collaborative training and semi-supervised liver CT image segmentation method based on classification.(1)The main work of semi-supervised liver CT image segmentation method based on deep collaborative training is as follows: The collaborative training method in semi-supervised learning just cross-joins the pseudo-labels generated by the two classifiers on the unlabeled data to the training set,which will introduce a large error and the two classifiers generate two results make it difficult to determine the final result.This paper proposes semi-supervised segmentation method based on deep collaborative training.This method through two steps of rough selection and fine selection,the pseudo-labels with higher confidence are selectedand added to the training set,which reduces the errors brought by the pseudo-labels.The average-cut selection criterion determines the final segmentation result,and applies the morphological opening operation to the post-processing of the segmentation result.Iterates the above process until the Dice and Jaccard similarity coefficients of the segmentation result of the verification set no longer increase.The experimental results show that compare with multiple fully supervised learning methods under the same training set,the Dice of the proposed method is 2.73 percentage points higher than that of U-Net,and the Jaccard is 3.02 percentage points higher.Compare with 2D V-Net Dice increase by 0.78 percentage points,Jaccard increase by 1.09 percentage points.When the comparison data included only images of the liver region in the test set,the Dice of the proposed method is 12.55 percentage points higher than that of U-Net,and the Jaccard is 14.15 percentage points higher.Compare with 2D V-Net Dice increase by 3.7 percentage points,Jaccard increase by 4.89 percentage points.(2)The main work of semi-supervised liver CT image segmentation method based on classification is as follows: In the self-training method for semi-supervised learning,the initial training set includes a large number of images of liver-free regions.The segmentation of these images is of little significance for computer-aided diagnosis and will cause serious class imbalance problems.And it is prone to over segmentation,reduce the accuracy of the overall segmentation result.This paper proposes classification-based semi-supervised segmentation method,First use the classification network to classify the data set.The classified images with liver regions are used in the following segmentation tasks to effectively alleviate the problem of class imbalance.Then use the maxmin-cut selection criterion adds a pseudo-label with higher confidence to the training set,and retains the 3D maximum connectedregion as the post-processing of the segmentation result.The experimental results show that under the same data set,the Dice of the segmentation result after adding the classification step is increase by 2.33 percentage points,and the Jaccard is increase by 2.22 percentage points.The Dice of the proposed method increase by 4.66 percentage points,and the Jaccard increase by 4.53 percentage points.The above two methods respectively improve the collaborative training method and the self-training method in semi-supervised learning,and improve the accuracy of the segmentation results compare with the fully supervised learning under the same data set.