基于深度特征融合的内窥镜图像分割网络OA

Gastroscope examination is a medical diagnostic technique for early diagnosis and preventative treat-ment of gastric cancer through detection and resection of lesion mucosa.Deep learning method has great application po-tential in gastroscopy.There is currently a lack of publicly available early gastric cancer data sets for researchers to use,and existing endoscopic segmentation methods cannot efficiently identify gastroscopic lesions.To solve this prob-lem,a new dataset Colonoscopy is proposed in this study,including the segmentation task of early gastric cancer lesions and the multi-classification task from healthy to early cancer stage.An endoscopic image segmentation architecture based on deep feature fusion is also proposed.A pre-trained MTE(Mixed Transformer Encoder)is used as the main work,and a deep fusion feature pyramid decoder is used to achieve accurate lesion segmentation.The proposed method has better performance in Colonoscopy data set,and has better generalization ability in two large colonoscopy data sets including Kvasir-Seg and CVC-ClinicDB.