Research On Segmentation Technique Of Esophageal Cancer Based On Deep Learning

Esophageal cancer is a common malignant tumor in life,and there are a large number of new patients with esophageal cancer every year.Meanwhile,the incidence rate and mortality rate of esophageal cancer are ranked at the top,which has seriously affected human health.In this thesis,we use deep learning image segmentation technology to analyze medical images of esophageal cancer,which brings new opportunities to the clinical diagnosis of esophageal cancer.This thesis is mainly based on Unet network,and based on this,we improve the effectiveness of image segmentation by adding self-attention mechanism to better learn global and remote semantic information interaction.In this thesis,through the training and learning of Trans Unet and Swin Unet,a novel model based on Transformer structure,we effectively identify and classify images as well as extract target regions,reducing the workload of doctors while improving the confirmation rate of esophageal cancer segmentation.In order to solve the problems of low efficiency and poor accuracy of traditional manual segmentation of esophageal cancer images,this thesis studies and realizes the improvement of efficiency and accuracy of esophageal cancer segmentation technology based on deep learning.Firstly,through literature query and video learning,we understand the relevant medical knowledge and characteristics of esophageal cancer lesions,and use the labeling tool Labelme to manually label the esophageal cancer image data and process the format of labeled images.Under the control of other variables,all three models are randomly divided into training and testing sets according to 8:2.Due to the small amount of data in this thesis,methods such as data augmentation and increasing the number of iterations are chosen to avoid problems such as model overfitting.In this thesis,we compare and analyze the segmentation effects of three models,Unet,Trans Unet and Swin Unet,and find that the Trans Unet model has the highest segmentation accuracy(0.9076)and the Unet model has the lowest segmentation accuracy(0.8779),indicating that the Transformer encoder,which is based on the attention mechanism,can better capture the image in global information and long-distance dependencies,thus improving the segmentation accuracy.However,due to the small training samples,the two large models,Trans Unet and Swin Unet,do not have outstanding advantages over the Unet model in terms of segmentation accuracy.And because the Trans Unet model combines both convolutional network and Transformer for feature extraction and fusion,it has higher segmentation accuracy than the Swin Unet model.This thesis demonstrates the effectiveness and superiority of deep learning techniques introducing self-attention mechanism in esophageal cancer segmentation,and provides new ideas for esophageal cancer segmentation techniques.