Research On Intelligent Detection And Segmentation Of Early Gastric Cancer Lesions

Incidence and mortality of gastric cancer are increasing significantly worldwide and threaten human life and health seriously.Early detection and timely treatment of gastric cancer are of great chance to cure.In clinical practice,endoscopy is a common tool for the examination of various diseases of the digestive tract.However,the endoscopic diagnosis of early gastric cancer has a low detection rate,high missed detection rate,and high misdiagnosis rate.Deep learning has developed rapidly in recent years,computeraided diagnosis（CAD）methods have been proposed increasingly,aiming to provide reliable and timely supplements to endoscopic diagnosis,and further to improve the accuracy and efficiency of diagnosis.Therefore,this thesis studies the intelligent detection and segmentation of early gastric cancer lesions using deep learning.The main contents are divided into two parts shown as follows:（1）Research on video and image detection method of early gastric cancer based on YOLOv5.An improved YOLOv5 for early gastric cancer detection was developed.Firstly,coordinate attention mechanism is added to the feature extraction part of the network to strengthen the feature of early gastric cancer,and the weighted fusion of the extracted feature information at different scales is to increase the diversity and robustness of the features.Then,the cluster-nms algorithm is used to adjust the screening strategy of the bounding box to improve the accuracy of the improved YOLOv5 and reduce the training time.Finally,the network uses 4 detection layers for lesion detection and location.The results show that the precision,recall and m AP of the proposed method are 93.9%,72.1%and 77.6%,respectively,which are better than the original YOLOv5,and the detection speed on video is 26 FPS,,which meets the needs of real-time detection.（2）Research on image segmentation method of early gastric cancer based on U-Net.Inspired by the encoder-decoder structure of the U-Net,an improved UNet network for segmenting the area of early gastric cancer was developed.Firstly,the pre-activated residual unit is used as the basic block to build the network.The channel attention mechanism is incorporated into the encoder module to adaptively weight the extracted features and highlight useful feature representations.Finally,a global average upsample module is added to the decoder module to provide global context information which can be a guide for the low-level features and alleviate the poor recognition performance when low-level features are concatenated with corresponding high-level features.The proposed method achieves a dice similarity coefficient of 74.19%,which is superior to other contrast methods.The CAD methods developed in this thesis show good performance,and can provide doctors with reliable and timely objective basis.They are with great clinic application prospects.