Single-Image-Based Deep Learning For Interactive Endoscopic Lesions Segmentation

Gastrointestinal(GI)disease is one of the most common diseases and primarily examined by GI endoscopy.Image segmentation is a key task in GI disease diagnosis,which aims to accurately identify target organs,tissues,and lesions in endoscopy images and provide basic information for further diagnosis.Traditional segmentation methods do not require any training dataset,but it is difficult to provide sufficiently accurate results,while deep learning-based segmentation methods usually face the problem of lacking sufficient datas,which results in prediction results of adequately trained networks that are not robust for images from different patients.To overcome this challenge,a single-imagebased deep learning method is proposed,with specific research efforts including:(1)A single image based dataset generation algorithm(Rendering From Single Image Lesion,RFSIL)is designed,which includes two innovations: The first is an interactive coarse labeling method based on clinical medical expertise,which uses Region of interest(ROI)selection and sampling of lesion regions to achieve coarse-grained image label acquisition.Second,we designed a simple graphics-based image rendering method to generate a diverse training set from a single input image by cutting and pasting the sampled focal areas.(2)A novel convolutional neural network framework: EUnet(edge-enhanced UNet)is introduced,and an edge decoder is added to U-Net,including an edge detection module responsible for edge extraction and an edge attention module for adaptive edge feature fusion to achieve accurate recognition of the edges of lesions to be segmented.an edge enhancement module with a new composite loss function.Meanwhile,a new combine loss function with edge enhancement module is designed to realize the cooperative work between the edge decoder and the segmentation decoder,so as to improve the edge performance of the segmentation results.The simulation results show that the improved model’s mean Dice coefficient(m Dice)improves by 3.9% compared to U-Net.(3)For better research,a precisely labeled Early Esophageal Cancer(EEC)dataset,EEC-2022,the first publicly available dataset for segmentation of EEC,was produced.Combining the above improvements,the proposed You-Only-Have-One(YOHO)achieves m Dice of 0.940 and 0.924 on two publicly available polyp datasets,CVCClinic DB and Kvasir,respectively,and 0.877 on early esophageal cancer dataset(EEC-2022)created by ourselves.The results show that,compared with the benchmark model,YOHO performs outstanding on endoscopy images,especially early esophageal cancer images,and achieves a significant performance improvement with lower network complexity,reaching the state-of-the-art(SOTA).