Research On Detection And Segmentation Automatically Of Lesion In Intestine By GAN

The intestine is the main organ for the body to absorb nutrients and energy,so once the disease occurs,it will have a serious impact on human health.Crohn is one of the intestinal diseases,mainly manifested as the onset of and the persistence of ulcerations of the intestinal wall.The diagnosis of Crohn requires the visual examination of the digestive tract by an expert who detects and characterizes the lesions.The one of effective way to get the intestinal image is to use videocapsule endoscopic that is a small capsule carrying a video camera.However,the data returned is huge because videocapsule endoscopic will stay the patient intestine for 6 to 7 hours.Although detection network based on CNN can separate lesion images,the expert still need to annotate the position of lesion respectively which also is a huge workload.Traditional segmentation network seems to be possible solution that can realize the identification and segmentation of target regions in images,so as to achieve locating the lesion.However,for training such network,it requires a lot of cost to manually mark the target regions in the dataset,which increases the difficulty of research so that it is difficult to apply in practice.To solve the problem of difficult training and high cost of marking dataset,I introduced a network achieving lesions detection and location automatically by unsupervised learning.The proposed method for predicting target area is made up by a binary classified CNN and an unsupervised GAN.The core idea is to use GAN’s application in the field of image inpainting that inpaints the image for erasing the target needed to be detected,and then combine with the binary classified CNN to predict the location of the target according to the change in the probability of the existence of lesions before and after inpainting.In this approach,GAN has the unsupervised characteristic during the training process,so that the dataset for training is not required to be marked with target location,which greatly reduces the cost of manually labeling the training set.In the experiment of detection and location of Crohn’s lesion,the method has achieved considerable and expected results.In order to realize the detection and location of lesion,the method can be divided into three parts:(1)Lesion detection part,which is made up by a binary classified CNN trained with healthy and diseased dataset to detect the presence of lesions in the image;(2)Lesion inpainting part,which is made up by GAN-based inpainting network trained by healthy intestinal images to achieve reconstructing missing content in the intestinal image into healthy intestinal wall;(3)Heat Map generator,which combines detection and inpainting module to generate and update the Heat Map predicting the location of the lesion.The overall system steps are that: The intestinal image first enters the lesion detection module,where the image detected as a lesion will be covered by the sliding mask window and sent to the leison inpainting part.At the same time,the heat map of the lesion is initialized.The image after inpainting will be sent to the lesion detection part again for calculating the new lesion probability.Later,the Heat Map generator will update the Heat Map according to the change of the probability of presence of lesions after each inpainting step and the current position of the sliding mask window.Then,the sliding mask window moves to the next step and repeats the steps of inpainting,detection,and Heat Map update until the sliding mask window finish scanning the entire image.This paper proposed a method of lesion detection and segmentation based on GAN,which achieved locating the target area by unsupervised training.Compared with the traditional image segmentation network,it significantly reduces the cost of manually labeling dataset during the training process.In the experiment of detection and location Crohn’s disease,the Heat Map show similar position and shape as ground truth segmentation of lesion.In the diagnosis,it can assist experts to locate the disease,thereby reducing the expert’s workload.In addition,in the research of the inpainting part,I proposed a type of inpainting GAN based on partial convolution that has achieved better performance in the intestinal environment than the existing advanced inpainting network.