Feature Extraction And Recognition Of Common Lesions Associated With Chronic Gastritis

Early detection, diagnosis and treatment of gastric cancer can significantly increase the survival rate and improve the quality of patients’daily life. Studies have shown that the majority of patients with gastric cancer showed symptoms of chronic gastritis in earlier time, thus providing an effective computer-aided diagnosis and decision support methods is quite important and meaningful for early diagnosis of gastric cancer.This paper presents a method using comprehensive features for feature extraction and recognition of endoscopic images. This method aims at recognizing three typical lesions in endoscopic images. It provides decision supports for the early diagnosis of gastric cancer during gastroscopy procedure. Two specific methods are utilized to classify these lesions. First, we combined color wavelet covariance texture (CWC), curvelet and local binary pattern texture (CLBP) which showed outstanding performances in feature extraction of endoscopic images. After that, support vector machine (SVM) and multilayer perceptron (MLP) were applied as classifiers for image classification. The second method utilized sequential floating forward selection (SFFS) to select features from lesions. Each lesion had its own SVM model as the classifier, and each classifier was only responsible for one type of gastric lesions, which is more accurate than one SVM model responsible for all the three gastric lesions. Specific steps are as follows:1) Comprehensive feature extractions of three lesions. Major image features extracted in this paper are texture features and color features. SVM and MLP were chosen as classifiers, which had broader applications in image classification tasks. For each lesion, different color spaces were applied. Texture features were calculated on single color space channel. Curvelet transform combined with local binary pattern (LBP) formed CLBP feature, and wavelet transform combined with gray-level co-occurrence matrix (GLCM) formed CWC feature. Finally, single channels were combined together to form a comprehensive feature, SVM and MLP were utilized to classify different lesions in different color spaces.2) SFFS for feature selection in lesions. This paper combined SFFS with SVM to extract image features which can better depict three lesions. Four color spaces were chosen, including RGB, YCbCr, K-L and HSI space. Each color space was divided into three single channels, and texture feature was calculated on single channels. Firstly, wavelet transform and curvelet transform were applied, thus producing lots of images. This method calculated statistical features of the histogram in those images, and five statistical features were adopted, including the maximum, the minimum, the average, the dominant and the extension. All these features were combined together to form a feature collection. Finally, for each lesion, SFFS and SVM were used to select subsets from the collection, and the model was chosen from all the SVM models of the lesions.3) System and Experimental results. Based on the comprehensive features and SFFS, my work, accomplished feature extraction and recognition of erosion, ulcer and atrophy. Based on feature extraction, a computer-aided decision support system (CADSS) which studies three lesions is constructed. After doing these, the feature base has been established with features of erosion, atrophy and ulcer. The feature base has been evaluated by clinical endoscopic images, and the result shows this system can basically recognize erosion, ulcer and atrophy.