Colonoscopy Image Classification System Based On Multi-Feature Fusion And Saliency Detection

In recent years,colorectal cancer has become one of the cancers with high morbidity and mortality worldwide.Colorectal polyps have a potential risk of deteriorating into tumors.Early detection and removal can significantly improve survival rate.The improvement of people's awareness of prevention have contributed to the growing number of colonoscopies performed by hospitals but the detection of polyps from the generated mass images has gradually become a difficult problem for doctors to diagnose.Missing polyps also become one of the major risks of colonoscopy.Thus,research on intelligent classification and computer-aided diagnosis for colonoscopy images has great significance.In this thesis,a classification system of colonoscopy images based on multi-feature fusion and saliency detection is proposed by the characteristics of intestinal polyps.It is designed to detect polyps from massive images and to provide a computer-aided diagnosis for doctors.Classification result is limited in the models that only rely on feature analysis,since huge similarity information can be found between positive and negative samples.Therefore,digital image processing,machine learning algorithm and significance detection are combined and applied to the classification of colonoscopy images.The main contributions are summarized as follows:1.A real-time and efficient pre-processing system is designed to construct intestinal image database,containing the region of interest extraction,spot detection and inpainting and poor image removal to avoid interference factors as far as possible,especially the additional boundary generated by the highlight spots.2.A classification framework for colonoscopy images based on multi-feature fusion and dictionary learning is designed.Feature extraction is the core of machine learning algorithm and fused features are mainly based on color and texture in this thesis.The complementarity among features improves discriminability of samples.Dictionary learning based training is used to learn features and final feature vector is formed using sparse coding,which is trained and classified in the support vector machine.There is no storage burden of sparse training data which can simplify the learning tasks as well.3.A saliency detection model for intestinal polyps is proposed,which is mainly based on manifold ranking algorithm and is with the help of polyp center priori.The shape feature of intestinal polyps is helpful for difference operator and edge detection to fit polyp center.Then combined with multi-scale manifold ranking,the constructed saliency map can effectively detect polyps and ignore the background area.At this time,the extracted features are more focused on polyps,which can solve the similarity problem between positive and negative samples.4.The simulation of the proposed colonoscopy image classification system is tested by the constructed database to verify the effectiveness.