Application Research Of Intestinal Tumor Detection Algorithm Based On Convolutional Neural Network

The intestinal tumor is a common intestinal disease,including polyps and adenomas.Polyps are the early symptoms of intestinal diseases.Adenomas are relatively serious intestinal tumors,which are likely to become cancerous.The endoscopic intestinal examination is a kind of detection method for intestinal diseases.Intestinal images are collected by using the intestinal endoscope to go deep into the intestinal tract.Through the judgment of the images within the scope of endoscopic vision,the lesions are found to the maximum extent,so as to determine the patient’s condition.Endoscopic intestinal detection has the advantages of high detection efficiency and low cost,and can basically detect all intestinal areas.It has been adopted by major hospitals around the world.Accurate detection and basic classification of tumors(polyps and adenomas)are important targets for detection of intestinal tumors.Currently,detection methods of tumors mainly rely on human eye observation and judgment based on doctors’ experience.This method has low accuracy,low efficiency and relies heavily on doctors’ comprehensive ability.Because of the improvement of computer hardware computing power and the breakthroughs in relevant theories,deep learning has made great breakthroughs in the field of computer vision.In order to accurately detect tumors and accurately identify polyps and adenomas,this paper constructs the intestinal polyps and adenomas data set under white light,have more applicability than NBI light data sets,the improvement of the current mainstream network Faster-RCNN,and the optimization,enhances the detection accuracy of intestinal cancer,so as to promote the practical application of tumor detection algorithm.For the detection of intestinal tumors,the main contents of this paper include the following aspects:1.Two kinds of intestinal tumor data sets under white light were constructed for model training,validation,and testing;one was used to train the single-class tumor detection model(polyps and adenomas were labeled as one type),and the other was used to train the two-class detection model(polyps and adenomas were labeled as two types).A large number of images of intestinal tumors under the white light of intestinal endoscopy were collected.With the assistance of doctors,accurate labeling of tumor images was achieved by analyzing corresponding pathological reports.The label information is the location frame coordinate and category of the lesion.2.Using transfer learning,neural network models suitable for single-class and two-class detection of the tumor were trained on small sample training sets.3.Two improvements were made to the target detection network Faster RCNN to improve the accuracy of intestinal tumor detection.Based on Faster RCNN,the feature extraction ability of the network was improved by improving the trunk network,and two detection models with better detection effects were constructed by combining RPN network,ROI Pooling and border regression.They were named A-faster RCNN and B-faster RCNN,respectively.4.According to the characteristics of B-faster RCNN itself,it was improved and named C-faster RCNN,so that the network had higher accuracy under small batch size.Through group normalization,the accuracy of the network will not be changed due to the batch size change,thus improving the target detection accuracy of the network under the small batch size.Experimental results show that the intestinal tumor data set constructed in this paper can train a neural network model suitable for detection under white light,which is more universally applicable than the data set under the NBI light source widely used at present.The improved C-faster RCNN in this paper has a detection accuracy of 99.44% on the single-class data set and 91.24% on the two-class data set.Compared with other target detection networks and conventional image recognition methods,the detection accuracy has been significantly improved.