Study Of Lesion Detection Metnoas In Gastronlntestinal Image

In the diagnosis of gastrointestinal diseases,endoscopic detection technology produces a large number of gastrointestinal images,and it is time-consuming and laborious to identify the lesions manually.Moreover,problems such as the concealment of pathological features and the shaking of the electronic lens shooting process may cause missed diagnosis and misdiagnosis.Therefore,it is of great practical significance to study the automatic detection method of gastrointestinal image lesions.In this paper,we employ multi-pattern recognition algorithm and deep learning algorithm to realize automatic detection and classification of three common lesions in the gastrointestinal tract,i.e.esophageal cancer,polyp and bleeding.The main works are summarized as follows:1.This paper reviewed and analyzed the detection methods of these three lesions in gastrointestinal images,and the feature extraction method and target detection algorithm of the three lesions are mainly studied.2.A method of lesion recognition based on manual feature extraction and multi-pattern recognition is proposed.First,the HSV color space and the gray gradient co-occurrence matrix are combined as the extracted features.The feature vector containing 24 values which can reflect the color and texture difference of the lesion.Then,the feature is input into a random forest classifier for classification and recognition.Experimental results show that,By combining color features and texture features with random forest classifier,the recognition rate can reach 87.5%,which is 7.5%and 34.5%higher than the results of support vector machine and the extreme learning machine respectively.3.The deep learning algorithm SSD(Single Shot Multibox Detector)was used to detect lesions in gastrointestinal images.The SSD algorithm can directly return the category of the lesion at that location and its confidence at each location of the input image.The basic network is vgg-16,and the full connection layer is changed into the convolution layer,and four progressively decreasing convolution layers are added as the new network structure.In the predictive network,the same convolutional layer uses anchor prediction targets of different scales to solve multi-scale target detection problems,while taking into account speed and accuracy.The experimental results show that the average accuracy of the SSD(mean Average Precision,mAP)can reach 72.83%,which is 0.66%higher than the Faster-RCNN.The single frame image detection speed is about 4 times that of the Faster-RCNN algorithm.The problem of repeatedly marked lesions in the detection results of SSD algorithm was analyzed,and the feature mapping association between different layers in the prediction network was added,and the detection effect was significantly improved after the structural improvement.