The Study Of Automatic Recognition And Segmentaion Methods Of Gastric Lesion On ME-NBI Endoscopic Images

Gastric cancer(GC)is one of the common malignant tumors of the digestive tract and plays an important role in malignant tumors worldwide.In 2018,gastric cancer became the fifth most common cancer and the second leading cause of cancer deaths in the world.Therefore,the effective diagnosis and timely treatment of GC are urgently needed.Magnifying narrow-band imaging(ME-NBI)endoscopy is the most effective and direct method for screening and diagnosing stomach diseases.It is significantly important for early detection of gastric cancer-related lesions,timely treatment,reduction of the incidence and mortality of GC,and improvement of cure rate and 5-year survival rate.In order to overcome the long learning curve and learning difficulty of ME-NBI endoscopic diagnosis technology,the subjectiveness,the low efficiency as well as false detection and missed detection,a computer-aided methods for ME-NBI endoscopic image diagnosis is needed,which can help doctors improve work efficiency,improve inspection accuracy,and decrease false positives and missed detection rates.This research mainly studies the automatic segmentation and recognition methods of gastric cancer-related lesions in ME-NBI images,including the following three parts:(1)A novel Hue-texture-embedded active contour segmentation model was proposed for segmenting gastric cancer-related lesions in ME-NBI images to eliminate the interference of non-lesional areas,help doctors accurately diagnose lesions and collect classification samples in ME-NBI images.Therefore,this part proposed a global hue energy functional and local texture energy functional to represent the global color and local texture characteristics of ME-NBI images,respectively.These two energy functions were integrated into the active contour model to construct a Hue-texture-embedded active contour model for segmenting lesion regions of ME-NBI images using the level set.The experimental results showed that the proposed segmentation model and energy functionals can effectively realize the segmentation of Four different types of lesions in ME-NBI images with blurred boundaries.(2)ME-NBI gastric cancer image classification is one of the main causes of misdiagnosis of doctors.In order to help doctors improve the recognition rate and provide an objective diagnosis basis for doctors,this part used deep transfer learning to automatically classify ME-NBI images with small sample size and intra-class variation.In this part,the ME-NBI images were successfully divided into three categories: chronic gastritis,low-grade tumor,and GC by using the ME-NBI image features transferred from the deep convolutional neural network(DCNN).The experimental results showed that deep transfer learning can well perform the classification of gastric diseases in ME-NBI images with small sample size and intra-class variation and the classification performance of deep transfer learning outperformed traditional texture feature extraction methods.(3)This part proposed an interpretive fusion-semantic Chain-bridge(FSCB)transfer neural network model for rating GC in ME-NBI images with a small number of labeled samples as well as inter-class similarity,in order to help doctors to identify GC staging and give precise treatment and surgical methods.In order to solve the problem of small sample size and inter-class similarity,this part simultaneously applied deep transfer learning and expert semantic knowledge combined with a deep neural network.Firstly,this part designed a multi-step Chain-bridge transfer learning method and obtained a pre-training DCNN model to solve the problem of small sample size.Secondly,based on the texture feature extraction method,the abstract expert semantic diagnosis knowledge was quantitatively encoded into a digital image(semantic graph),which was used to settle the problem that abstract semantic knowledge cannot be input into the DCNN model.Thirdly,we separately input different semantic images and original RGB images into the pre-trained model obtained by Chain-bridge transfer learning and obtained different feature representations.Finally,a deep neural network was designed to fuse original image features and semantic features,learn features and train these features;then an FSCB transfer neural network model was obtained,which was used to classify GC staging of ME-NBI images.The experimental results show that the FSCB transfer neural network model can successfully classify GC staying of ME-NBI images with mall sample size and inter-class similarity,and explain the prediction results of GC.