Research On Detection Of Digestive Artifacts And Lesions In Endoscopic Images Based On Deep Learning

Endoscopy is the most basic and direct way to examine gastrointestinal lesions clinically.However,the images of the digestive tract collected by endoscopes contain a lot of artifacts,such as specularity,motion blur,bubbles,etc.The existence of these artifacts is not conducive to the diagnosis and treatment of the doctor.Moreover,when developing algorithms of image frame quality assessment,these artifacts will be the main considerations.Therefore,the accurate detection of artifacts in gastrointestinal endoscopy image is very important for the development of automatic quality assessment algorithms and computer-aided diagnostic tools of gastrointestinal endoscopy images.In addition,during the artificial diagnosis of gastrointestinal lesions,the huge amount of data of the endoscopic image and the diversity of the appearance of the lesions increase the difficulty of the doctor's diagnosis,and is likely to produce missed diagnosis and misdiagnosis.Therefore,the development of an automatic segmentation algorithm for the digestive tract lesion can help reduce the burden on doctors by assisting doctors in diagnosis and treatment to a certain extent,and help reduce the rate of missed diagnosis and misdiagnosis,which has great practical significance.To this end,this paper studies the methods in the detection of artifacts and the segmentation of lesions in endoscopic images of the digestive tract.The main research contents are as follows:(1)Research on the detection of artifacts in digestive tract endoscopy images based on deep learningAn object detection network based on Cascade RCNN(Cascade Region with Convolution Neural Network)was developed,and a cascaded detector was used to detect multi-scale targets.Firstly,the backbone network Resnet50 is combined with a feature pyramid network for extracting multi-scale features to enhance feature extraction capabilities and richer information.Then the Cascade RCNN is used to detect seven artifacts such as specularity,motion blur,bubbles,Saturation,Contrast,instrument and other imaging artifacts.In order to overcome the problem of adjacent and overlapping multiple artifacts,the soft non-maximum suppression algorithm is used to adjust the screening strategy of the detection boxes.The focal loss function is used to overcome the problem of imbalanced distribution of categories and impose differentpenalties on samples of different classification difficulty levels.The proposed method was cross-validated on a dataset consisting of 2347 endoscopic images,and a mean average precision(mAP)of 0.5412 and an intersection over union(IoU)of 0.4675 were obtained.Compared with other methods,the proposed method has better performance.(2)Research on the segmentation of gastroscopic early esophageal cancer based on deep learningInspired by the encode-decode framework of the Deeplabv3+ network,a new FC-DeepLab for the diagnosis of early esophageal cancer was proposed.The backbone network of the encoder is a simplification of that of Deeplabv3+,and the fully connected attention module is added,and then the atrous spatial pyramid pooling structure is used to further extract and integrate features at different dilated rates.The decoder restores the image resolution by bilinear interpolation,performs pixel-level classification,and obtains segmentation results.In addition,changing the activation function to PReLU can activate those neurons with negative input.Experiments show that the proposed method can improve the performance of the segmentation of early esophageal cancer.The loss function is the focal tversky,which can achieve a better balance between the precision and the recall,and reduce the missed detection rate.The performance on 3190 clinical gastroscopic images shows that the precision,recall and Dice coefficients of the proposed method in this paper are 0.7610,0.7814 and 0.7711,respectively,which achieves the expected effect.