Research On Segmentation Algorithms Of Early Gastroesophageal Carcinoma Based On Gastroscopic Images

Gastroesophageal cancer is a high-incidence cancer in humans and has a very high mortality rate.Timely interventional therapy in the early stage of gastroesophageal cancer can greatly improve the 5-year survival rate of patients.At present,early gastroesophageal cancer mainly relies on endoscopy,which is heavily dependent on the personal experience and professional level of doctors.In addition,the pathological characteristics of early gastroesophageal cancer are not significantly different from other upper gastrointestinal diseases,which may easily lead to missed diagnosis and misdiagnosis..Therefore,it is necessary to develop a Computer Aided Diagnosis（CAD）system for early gastroesophageal cancer.In this paper,deep learning（DL）is used as a technical solution to develop CAD for early esophageal cancer segmentation.Based on the collected endoscopic images of early gastroesophageal cancer,the segmentation algorithm for early gastroesophageal cancer based on DL technology is studied.The main research contents of this paper are as follows:Early gastroesophageal cancer segmentation algorithm based on improved UNet convolutional neural network（Convolutional Neural Network,CNN）.Aiming at the shortcomings that CNN can effectively extract local features,but cannot effectively extract global features,this paper proposes a Link-Context Module（LC-Module）to help CNN-based algorithm models effectively extract larger-scale features.A new algorithm network LC-Net is proposed.LC-Net can effectively build context global dependencies,enrich the feature maps extracted by the network,and enhance the semantic segmentation ability of the network.4322 early gastric cancer images and 2689 early esophageal cancer images were used to train and verify the LC-Net segmentation model proposed in this paper.The Dice similarity coefficients for segmenting early gastric cancer and early esophageal cancer reached 76.37% and 77.05%,respectively.compared to the performance of other UNet improved networks.Research on segmentation method of early gastroesophageal cancer based on Transformer technology.Aiming at the disadvantage that CNN cannot effectively extract global features,in order to fundamentally solve this problem,this paper introduces the Transformer method based on the Self-Attention（SA）mechanism,and proposes a new Deep PVT network for segmentation of early gastroesophageal cancer.Using the Transformer method as the backbone network,the ASPP module in Deeplabv3+ is improved to form a new decoder module,which enhances the fusion of shallow features and deep features,thereby enhancing the segmentation ability of the Deep PVT network.Using 4322 early gastric cancer images and 2689 early esophageal cancer images to train and verify the Deep PVT proposed in this paper,the Dice similarity coefficients for segmenting early gastric cancer and early esophageal cancer reach 80.5% and 79.8%,respectively,which are better than other CNN networks.