Research On Key Technology For Computer Aided Diagnosis Of Early Esophageal Cancer Based On Semantic Segmentation

Esophageal cancer is one of the most common malignant tumors that occurs in the esophageal mucosa,with about 500,000 deaths per year.Studies have confirmed that early detection of lesions can effectively reduce mortality.Endoscopy is a common method for early detection of esophageal cancer.However,with the increase of imaging rate,the burden of doctors’ reading is obviously aggravated,and a doctor’s diagnosis has certain subjectivity.Therefore,automatic identification and accurate localization of early esophageal cancer by computer is of great significance to improve the accuracy of diagnosis and reduce missed diagnosis and misdiagnosis.At present,the research on the computer-aided diagnosis of early esophageal cancer mainly focuses on white light endoscopic images.Because the mucosal morphology of esophageal cancer is small and easy to be neglected,conventional ordinary white endoscopy is easy to miss diagnosis and misdiagnosis.This article mainly provides a computer-aided diagnosis for three key steps in the endoscopy of esophageal cancer: NBI images,IPCL images,and iodine-stained images,in order to assist doctors to improve the diagnostic accuracy of esophageal cancer.This paper focuses on the key steps of computer-aided diagnosis of early esophageal cancer and semantic segmentation technology to improve the accuracy and efficiency of computer-aided diagnosis.The main contributions and innovations are as follows:1.A new M-DeepLab network for computer-aided diagnosis of early esophageal cancer was proposed.The network combines all the features of the existing semantic segmentation networks,including encoder,decoder,feature fusion,and pixel classification.The encoder encodes the input image to extract features;the decoder decodes the extracted features and restores the image semantic information.Feature fusion is performed by adding an atrous spatial pyramid pooling module to extract features of different scales for fusion,and finally the classification of image pixel level is completed.2.The proposed M-DeepLab network was applied in three key steps in the endoscopy of esophageal cancer: NBI images,IPCL images in magnified endoscopy mode,and the iodine-stained images for the secondary diagnosis,respectively.On this basis,we constructed three datasets for model training,namely NBI dataset,NBI-ME dataset,and iodine-stained image dataset.Because the model training requires a large amount of data and medical images are usually small datasets,this paper proposed a data augmentation strategy for rotation,flipping and scaling,according to the characteristics of esophageal cancer images.3.The network was tested in three datasets,and the experimental results show that the overall accuracy of the NBI image is 97.15%,and the Mean Intersection over Union(MIoU)is 91.05%.In the five-category task of IPCL images,the overall accuracy reaches 94.80%,and MIoU reaches 89.90%.For the two-class iodine-stained image,the overall accuracy ups to 97.31%,MIoU reaches 92.09%.Moreover,it takes only 0.05 s to judge an iodine-stained image by the M-DeepLab model.The M-DeepLab model exhibits good performance both in accuracy and speed for early esophageal cancerous lesions identification,comparable to the experienced clinicians.As assistance for the clinicians,the proposed model could possibly increase the diagnosis rate and decrease the misdiagnosis of early esophageal cancer.