Research On Intelligent Detection And Recognition Of Small Intestinal Lesions Based On Wireless Capsule Endoscopy Images

In recent years,wireless capsule endoscopy(WCE)has been widely used in digestive tract examination because of its advantages of safety,painless,noninvasive,as well as being able to detect the whole small intestine which is difficult to reach by traditional endoscopy.However,for each examination,WCE will produce 50,000 to 80,000 images.Doctors need to spend a lot of time to review all images for diagnosis,resulting in waste of medical resources.Since the pathological images is in small proportion,it is easy to cause false and/or missed detection with the naked eye.Therefore,it is of great significance to realize intelligent detection and recognition of small intestinal lesions in WCE images.Several researches have been performed in the field of intelligent detection and recognition of small intestinal lesions in WCE images,most of them use traditional machine learning methods to detect single lesions.However,the traditional machining learning methods need to extract features manually,which limited the detection efficiency and accuracy.Taking advantages of automatic learning of image features and updating parameters by back propagation,deep learning has been widely used in WCE image lesion recognition studies,but most of studies were aimed at single lesion detection and recognition,few studies have focused on the detection and recognition of multiple lesions.Therefore,this study proposed a novel method for intelligent detection and recognition of small intestinal lesions in WCE images based on deep learning and transfer learning.The proposed method was carried out by using the deep convolutional neural networks(DCNN)and R-CNN(Region-CNN)methods.The main research contents are as follows:(1)Research on pyloric and ileocecal valve recognitionAutomatically recognizing the location of pylorus(the start of small intestine)and ileocecal valve(the end of small intestine)can help doctors to find the small intestine area to detect its lesions quickly and accurately.Through a combination of Res Nets and transfer learning,a model for detecting small intestine images was established.And then,with the proposed positioning algorithm,this study achieved to locate the pyloric and ileocecal valve accurately.The results show that the absolute deviation degree error of the two positioning was less than 0.5%,or even no error.(2)Research on intelligent detection for small intestinal lesionsThis study was aimed to achieve intelligent detection of small intestinal lesions,including hemorrhage,ulcer,erosion and polyp,based on Dense Nets and transfer learning.The experimental results demonstrated the proposed method could detect the lesions with high accuracy,sepcificity and sensitivity.And the ROC(Receiver Operating Characteristic)curve also proved the superior model performance on lesion detection with an AUC(Area Under ROC Curve)of 0.9985.(3)Research on small intestinal polyp lesion localizationThis study was to realize the localization of small intestinal polyps by combning deep learning and transfer learning.Comparing with Faster R-CNN and Mask R-CNN,the overall performance of Mask Scoring R-CNN(MS R-CNN)model with Res Net101-FPN(Feature Pyramid Networks)in localizing polyp was the optimal.Furthermore,an ensemble learning method based on output results of MS-RCNN with different feature extractor was proposed to further improve the performance on polyp localization.The evaluation metrics,including recall,precision,Jaccard index and Dice score was performed to demonstrate the optimal capacity of emsemble model,as well PR(Precision-Recall)curve and FROC(Free-response Receiver Operating Characteristic)curve.The experimental result showed that the AP value of the optimal model was up to 0.8411.(4)Research on the development of small intestinal lesions intelligent detection systemAn intelligent lesion detection prototype system was developed based on the pyloric and ileocecal valve recognition model and lesions intelligent detection model,which will be benefit for computer-aided diagnosis.Combing DCNN,R-CNN and transfer learning,this study realized the intelligent detection and recognition of small intestinal lesions.The promising experimental results showed that the trained model has higher detection and recognition accuracy.Based on the existing model,an intelligent detection software system was developed to realize the computer-aided diagnosis of WCE,which is beneficial to promoting the clinical application of WCE.