Detection Of Small Intestinal Tract Space-occupying Lesions Based On Feature Fusion In WCE Sequence

In recent years,Wireless Capsule Endoscopy(WCE)has overcome the shortcomings of traditional digestive endoscopy for small intestinal diseases.With its painless and noninvasive characteristics,WCE has been widely used in clinical upper gastrointestinal tract,especially in small intestinal tract detection,and has become the preferred equipment for digestive tract detection.However,the large amount of image data captured by WCE makes the clinical film-reading work a huge task,and it is easy to cause missed and false detection of lesions.To solve this problem,it is of great clinical significance and application value to develop an algorithm model for automatic detection of small intestinal tract lesions from WCE image sequences.In this study,we focused on the detection of small intestinal tract space-occupying lesions,and proposed a small intestinal tract space-occupying detection method for WCE image sequences based on two-stage deep learning network.The two-stage deep learning network proposed in this study is divided into spaceoccupying lesion detection network(stage 1)and space-occupying lesion identification network(stage 2).The two cascades constitute the overall detection framework of this paper.Among them,based on the traditional Faster R-CNN network,the feature extraction module and detection module are improved,and the parameters of anchor points in the regional recommendation module are adjusted to complete the screening and detection of small intestinal tract space-occupying lesions in WCE image sequence.The experimental results show that the proposed improved network is superior to the traditional Faster R-CNN network in extracting the proposed region of suspected lesions,with sensitivity of 98.37 %,false positive of 6.59 % and accuracy of 93.42 %,while the sensitivity,false positive and accuracy of 98.28 %,6.3 % and 93.71 % in the detection results of space-occupied images.Secondly,combined with residual structure and feature pyramid structure,a small intestinal tract space-occupying lesion identification network based on feature fusion is built to further reduce the false positive rate of space-occupying image in spaceoccupying lesion detection network.In order to integrate the performance of detection network and identification network,this study adopts cascade mechanism to improve the overall detection effect.The experimental results show that the sensitivity,false positive and accuracy of the proposed detection network model in this paper are98.81 %,7.43 % and 92.57 %,respectively,for the extraction of the proposed region for suspected lesions,while the sensitivity,false positive and accuracy of the detection results in the space-occupied image are 98.75 %,5.62 % and 94.39 %,respectively.Finally,in order to make the method of this paper get rapid application,combined with the Django development framework,completed a WEB framework based on WCE small intestinal tract space-occupying lesions automatic detection software.The software can receive WCE images submitted by clinicians and quickly return the automatic detection report.