The Research Of Medical Image Segmentation And Its Applications

Medical image segmentation is a basic task in the field of medical image processing,which aims to extract regions of interest(ROI)in medical images to assist doctors in diagnosing diseases.The use of deep learning to automatically segment organs,tissues or tumor regions from medical images has become a research hotspot.Convolutional neural network(CNN)is one of the most significant methods to achieve medical image segmentation.CNN-based network surrounding image segmentation firstly extracts the pixel features from the raw image and classifies the pixels according to the features;and then the classification results and category labels of pixels are employed to calculate the loss value.The parameters in the network are updated to reduce the loss value by the optimization algorithm,so that the classification results are consistent with the category labels.The segmentation network based on pixel classification fails to take advantage of the higher-order correlation between the categories of the center pixel and other pixels in the neighborhood,resulting in inaccurate boundary segmentation of the object,and the small object tends to be false or missed detection.It is very difficult to use the mathematical method to model the higher-order correlation of pixel categories in the neighborhood,and it is intractable to combine with CNN to achieve end-to-end training.In view of the above problems,this paper exploits neural network to learn the higher-order correlation between pixel categories from CNN segmentation results to improve the segmentation accuracy.The major works of this paper can be identified as follows: First,Energy-Based Generative Adversarial Network(EBGAN)is utilized to learn the higher-order correlation between pixel categories.EBGAN is composed of a generator and a discriminator.The CNN-based segmentation network and the auto-encoder are used as the generator and the discriminator respectively.The discriminator learns the higher-order correlation of pixel categories during the adversarial training with the generator.The advantage of the proposed method is that EBGAN designs the discriminator based on the energy theory to avoid vanishing gradient or exploding gradient,and the proposed method is tractable to train.Second,conditional variational auto-encoder(CVAE)is used to learn the higher-order correlation of pixel categories.This method uses CVAE to be a post-processing step to the output of the segmented network based on CNN,and CVAE is used to learn the higher-order correlation of pixel categories.The advantage of the proposed method is that the parameters in CVAE,and segmentation model can be optimized jointly to achieve end-to-end training.The above two methods are evaluated on the dataset of MICCAI 2017 Liver Tumor Segmentation(Li TS)Challenge and 3DIRCADb dataset,and the experimental results are verified with the medical image segmentation evaluation indexes.The results show that the proposed methods can effectively improve the segmentation accuracy of small volume objects and object boundaries.