Precancerous Disease Identification Method Based On Fusion Of Shallow And Deep Features

Gastric cancer is one of the most common malignant tumors in the digestive system.According to a report released by the National Cancer Center in 2018,gastric cancer is the second most common cancer in China,second only to lung cancer,and its mortality rate is also ranked third.The formation of gastric cancer needs to undergo the process of precancerous diseases,precancerous lesions and early gastric cancer.Precancerous diseases include gastric ulcer,gastric polyps,gastric erosion,and chronic gastritis.These diseases are accompanied by varying degrees of inflammation and atypical hyperplasia of the gastric mucosa,which can be converted into cancer.The timely diagnosis of precancerous diseases is conducive to controlling the disease and reducing the risk of cancer.The traditional diagnosis of precancerous diseases relies mainly on the manual identification of clinicians,which is time-consuming and requires the clinician to have rich work experience.In recent years,the rapid development of computer vision and the improvement of computing performance have provided reliable methods and opportunities for automatic identification of precancerous diseases.The gastroscopy image contains a large amount of digital information for clinical diagnosis,prediction and therapeutic effect prognosis,which can be high-throughput extracted by artificial intelligence methods such as the traditional machine learning methods and deep learning that has developed rapidly in recent years.Image recognition based on machine learning requires hand-designed shallow features of images,but shallow features only contain color,texture,spatial information,etc.,and lack deep semantic features.Its recognition accuracy is low and relies on well-designed medical image processing algorithms.Image recognition based on deep learning can automatically learn image features,but the size of lesions and the number of images directly affect the accuracy of recognition.Its classification results only depend on deep semantic features and ignore the shallow features of the images.In view of the above problems,a precancerous diseases classification model based on features fusion is proposed in this paper to automatically identity three types of precancerous diseases including gastric polyp,gastric ulcer and gastric erosion.A recognition system is constructed to provide clinicians with auxiliary diagnosis and support clinical decision making.Among them,features fusion refers to the fusion of two types of features that represent the shallow and deep information of the images.The shallow features of images include texture features,histogram features,and fractal dimension features.In this paper,75-dimensional manual shallow features are extracted,in which the histogram features are extracted by gray histogram;the texture features are extracted by Grey Level Co-occurrence Matrix,Grey Level Run Length Matrix and Grey Level Gradient Co-occurrence Matrix;fractal dimension features are extracted by the difference box counting.The deep features of images are represented by neurons on the fully connected layer in Convolutional Neural Network(CNN).In order to match the number of shallow features,75 full connected layer neurons are also used in this paper,and in series with the shallow features to form a new 150-dimensional quantitative feature.In order to eliminate the influence of the dimensions,each feature is normalized to[-1,1].Based on the above quantitative features that combine the shallow and deep information of the images,the precancerous disease images are classified and identified by using machine learning classifiers in this paper,such as Support Vector Machine(SVM),Random Forest(RF),BP neural network.The data used in this paper was obtained from the gastroscopic image information of 211 patients in a Chinese tertiary hospital.In order to ensure the balance of disease samples,380 samples of lesions in different image regions were selected from the above images for each type of disease.Among them,the lesion samples for shallow features extraction are marked based on the accurate contour;the lesion samples for deep features extraction are marked with a rectangular frame;the lesion samples are marked by the digestive doctor with many years of clinical experience in the tertiary hospitals.On this basis,the feature fusion model proposed in this paper is used to classify the lesion images.The experimental results show that the classification accuracy of the feature fusion model is up to 95.18%,which is much higher than the classification accuracy of 74.12%based on the manual feature model,and 2.64%higher than the classification accuracy of the convolutional neural network model of deep learning.Especially in the classification of gastric ulcer,the classification rate of the feature fusion model proposed in this paper is as high as 98.68%.Finally,the correlation between different types of features and precancerous disease classification results is analyzed by Lasso method in this paper.The first six high correlation features include five deep features and one shallow feature,which shows that the deep features and shallow features of the image have important significance for the classification of precancerous diseases.In summary,the precancerous disease classification model based on feature fusion is superior to the traditional machine learning classification method in the classification of precancerous diseases,and it also has a significant improvement in accuracy compared to the deep learning model.It indicates that the fusion of shallow and deep features can more fully characterize the information of the gastroscopic images.This method provides new ideas and technical means for the classification of precancerous diseases,and it also helps to provide clinicians with more effective diagnostic methods.