Monocular Depth Estimation In Gastrointestinal Endoscopy

Due to the complex diet composition and irregular eating habits of modern people,the incidence of gastrointestinal diseases is increasing year by year.If patients cannot get timely treatment,they will have repeated attacks for a long time,which can easily transform into cancer.Gastrointestinal endoscopy is an effective method for the diagnosis and treatment of gastrointestinal diseases.However,due to the limitations of monocular imaging of endoscopic images,it is difficult for doctors to objectively obtain the threedimensional spatial information of the video area,and it is easy to cause missed diagnosis in the process of gastrointestinal endoscopy.It is not easy to accurately grasp the depth of the lesion during biopsy and other operations,which brings unnecessary burden to the patient.This paper uses the sparse reconstruction data obtained by traditional methods to train a neural network to predict the dense depth map of monocular gastrointestinal endoscopy images,thereby presenting three-dimensional information,which brings convenience to doctors’ diagnosis and treatment.In the sparse point cloud reconstruction stage,this paper proposes an improved Structure From Motion(SFM)method based on region expansion matching.Compared with the traditional SFM method,the method in this paper can produce denser and more three-dimensional sparse reconstruction results.In the neural network training phase,this paper uses the results of sparse reconstruction to train the network model for dense depth prediction according to the Depth Consistency Loss(DCL)and Sparse Depth Loss(SDL).The experimental results prove that the dense depth map obtained in this paper has better three-dimensional structure information.This article uses the published laparoscopic video data and clinically collected gastrointestinal video data as experimental materials.The experiments of the traditional method are mainly carried out on stable and smooth laparoscopic data,and the clinical gastrointestinal data is used in the sparse reconstruction method and deep learning of this article.experiment.The qualitative and quantitative comparison proves the effectiveness of the method in this paper.