Colonoscopy Image Classification Algorithm Based On Feature Selection And Encoding

In recent years,the number of cancer cases in China has been increasing,the incidence and mortality of colorectal cancer keep significant growth.Studies have shown that many patients with colorectal cancer have a history of polyps and ulcerative colitis.Polyp usually grows slowly,and ulcer is a chronic condition.Early detection and treatment before the disease worsens can significantly improve the patient's clinical symptoms and reduce the incidence of colorectal cancer.Endoscopy is a common method of gastrointestinal examination and can detect intestinal conditions in real time,but it will generate a large number of images.It will take a lot of time and energy if only relying on the doctor,and the cases of missed diagnosis is inevitable during the long time work.Therefore,it is of great significance to study the computer-aided diagnosis system for automatic detection of colorectal diseases to improve the examination efficiency and accuracy.In this thesis,a colonoscopy image classification algorithm for automatic detection of polyps and ulcers is studied which based on feature selection and feature encoding.The main contributions are summarized as follows:1.The image preprocessing system was designed to deal with the interference signals such as black frame,bubble and specular in the original colonoscopy image.Firstly,the region of interest(ROI)is extracted from original images,and an unreferenced image quality evaluation index based on image segmentation and pixel mean is adopted,which achieve a more ideal results compared with the global image statistical method.Because of specular will produce false edges which similar to polyps or ulcers,a method of threshold-based detection and pixel filling is used to detect and repair the light spot,thus an effective colonoscopy image dataset is obtained for the design of image classification algorithms.2.According to the characteristics of intestinal image,a feature extraction method based on color and local texture description was proposed.Firstly,according to the method of histogram quantization,the image is quantized into several color subintervals in HSV color space.For each sub-interval pixel,the RGB difference between it and the adjacent pixels is calculated to represent the local changes of the image,and then the local color difference histogram(LCDH)feature is obtained.3.Considering that a single feature usually cannot describe the endoscope image accurately,and the series of several features will cause feature redundancy,thus an image classification framework based on multi-features fusion and feature selection is designed.Firstly,multiple color and texture features of each image are extracted.Then combined with the Constraint-Guided Sparse(CGS)feature selection algorithm to filter out more discriminative feature components from the feature pool.Finally,SVM classifier is adopted,and the classification accuracy achieves 93.44% in the binary classification experiment.4.To further improve the classification performance,an image classification framework based on improved bag of features model is proposed.The core step of the model is local feature extraction and feature encoding.In this thesis,the similarity between visual words and feature descriptors is introduced as the adaptive weight of codes based on the LLC coding method,thus a normalized variance local constrained linear coding(NVLLC)method is proposed.Then LCDH feature is adopted as local feature descriptor in the feature extraction step.And the feature learning strategy which only based on positive samples is also adopted to generate the codebook.The experimental results show that the classification accuracy achieves 97.50% in the binary classification experiment,and the recognition rate of polyp and ulceration in the multiclassification experiment achieves 92.50% and 95.50%,which verifies the effectiveness of the method.