Study On Esophageal Cancer Images Segmentation Algorithm Based On Deep Learning

Image segmentation in medicine refers to the method of identifying disaffected organs,tissues or related areas from patients’ diagnostic images,and describing the contour,volume and location of these areas.Currently,image segmentation is a prerequisite for disease diagnosis and treatment in clinical practice,which is usually performed manually by doctors with experience in diagnosis.With the increase of the rate of patient visits,the scale of medical image data is growing rapidly,which brings a lot of tedious work to doctors.Therefore,many researchers have studied the automatic segmentation method of medical image.Most of the early medical image segmentation methods were based on the classical methods such as pixel boundary,threshold and graph theory,which could not meet the requirements of clinical application in terms of segmentation performance and efficiency.Deep learning can learn the most representative deep features of images from the original image data in a data-driven way.These methods use many simple simulated biological neurons to connect and compose layers,and many layers are stacked into a neural network to achieve complex tasks.This algorithm can learn automatically from massive image data without designing complex image features manually.These algorithms solve tasks including understanding and classifying input feature vectors in large amounts of data.Although convolutional neural networks,a key technique for deep learning,were first introduced decades ago,it is only in recent years that these algorithms have demonstrated convincing success,elevating their status from an algorithmic idea to a general-purpose algorithm of artificial intelligence.And in recent years,complex deep learning algorithms have rarely been able to outperform humans in some tasks,but in some tasks,such as diagnosing lung nodules and breast nodules,they have performed better than humans.Based on the deep learning technique,this thesis studies how to improve the performance and applicability of the segmentation algorithm of Computed Tomography(CT)for esophageal cancer.In this thesis,the main work and contributions are as follows:(1)We improves an image generation algorithm of esophageal cancer.Combined with the feature of context information that can be collected by recurrent neural network,according to the real CT sequence images,the CT sequence images with correlation before and after images are synthesized on the basis of generating confrontation network,so as to achieve the purpose of data enhancement,and to train the subsequent segmentation neural network model.(2)Because the esophagus takes a small part in the whole thoracic section,in order to improve the efficiency of the discriminator,an additional local discriminator was added in cooperation with the main discriminator to identify the esophageal part on the basis of generating the antagonist network.The coarse segmentation module is designed in combination with the attention mechanism,which gives the network attention in the segmentation part.Through this module,the network can automatically learn which features need more attention from the data.Then,the U-shaped network deepened by the residual block was combined with the antagonist network discriminator to adjust the output results for fine segmentation.The coarse segmentation was responsible for distinguishing the tumor area from the background.The fine segmentation further separated the current tumor area from its adjacent soft tissues by the segmentation results of the former,and accurately segmenting the tumor boundary.(3)In terms of algorithm robustness analysis,three image segmentation network models,Cascaded NET,classic U-NET and Deeplab-v3,are trained respectively by using the conventional data enhancement algorithm and the training set generated under the data enhancement algorithm in this thesis.According to the evaluation indexes: DICE similarity index,classification error and volume error,compared with several other neural network algorithms used for image segmentation,it is verified that the robustness of the proposed algorithm is better than other algorithms in different tumor location features and noise affected environments.