The Research Of Medical Image Classification Methods Based On Compound Visual Words

Medical image classification technology is an important support to organize and store a large number of medical image data, and use the implied information to support clinical diagnosis and treatment. There are a lot of clinical applications of medical image classification. On the one hand, the global medical image classification is the key to improve the performance of medical images storage and medical image retrieval. On the other hand, local medical image classification is the main method to realize the local lesion recognition and classification. Previous studies using global features of medical images, cannot describe local details of medical images. Local features are used more and more in recent years. Bag-of-visual-words (BOVW) model has been widely referenced to realize medical image classification. However, the further research of both medical image visual word learning and visual words based medical image classification is necessary and valuable.This dissertation focuses on the field of visual words based medical image classification. The research is around the related problems in the process of classification. First, we define the compound visual words, and present a fundamental medical image classification method based on scale compound visual words. After that, we discuss in depth on medical image feature extraction, visual words learning and medical image classifier learning, in order to improve the medical image classification. The contributions of this dissertation are as follows:(1) A single feature vector is not sufficient to fuse multiple feature vectors extracted under different parameters, which affects the classification based on traditional bag-of-visual-words model. In order to improve the result of medical image classification, we defined the compound visual words, and present a medical image classification method based on scale compound visual words. First, Wavelet transform combining with spatial relationship is used to generate multiple features. A feature matrix is composed of the multiple features and used to represent visual words. The multiple features are named as scale compound features. Second, we use ensemble learning to learn and label the compound features composing with the multiple features. The labels of medical image categories are used to represent visual words. Finally, we use voting strategy on the results of ensemble learning to predict the categories of unseen medical images. Two public data sets are used to evaluate our method. The ImageCLEFMed database is used to test the global medical image classification performance, and the Emphysema database is used to test the local lesion recognition and classification performance. The experimental results show that our method can achieve higher accuracy.(2) The wide range of the intensity of medical images affects the representation effect of traditional features. We study in depth on medical image feature extraction, and present a medical image classification method based on contour compound visual words. This method extracts the contour of the texture in medical images. The main texture which are easy to be observed by human eyes are extracted, and the secondary texture which are hard to be observed are ignored. First, unsupervised method is used to cluster the pixels in the medical images. The points on the edges of the clusters are considered as suspected contour points. Second, clustering univalue segment assimilating nucleus are constructed and discriminant rules are used to remove the noise points to gernerate a set of contour points. The contour points are considered as contour visual words. Third, the descriptor of the contour points is defined. Contour compound feature is generated to represent medical images. Finally, ensemble learning is used to learn the contour visual words and the medical image classification is realized based on these visual words. Both ImageCLEFMed database and Emphysema database are used in our experiment. The experiment results show that the presented method achieved better performance.(3) Aimed at the lack of the judgment of the visual words effectiveness in the visual word learning, we study in depth on visual word learning. We present a medical image classification method based on compound visual word dictionary learning. On the one hand, the relationship between the visual words and their multiple features is considered in the process of visual word learning by using multiple instance learning. On the other hand, according to the results of the visual word learning, the "stop words" can be removed. First, the multiple instance model is constructed by using the compound visual words and the features of them. Second, the compound visual word dictionary learning is realized by a set of learners. The output of the dictionary learning is a set of rules represented by the learners. Finally, the visual words in unseen medical images are searched by using the visual word dictionary, and the medical image classification is realized based on the results of the search. Both ImageCLEFMed database and Emphysema database are used in our experiment. The experimental results show that the performances of the medical image classification are improved at different level by using the presented method. Especially the sensitivity and specificity of the emphysema slices classification are raised.(4) Aimed at the problem that a visual word may have multiple labels in the clinical enviorment, we study in depth on medical image classifier learning. We extend the multiple instance learning model and present a medical image classification method based on the class space of compound visual words. First, a class space is constructed, in which each axis represents a category of the medical images. Second, visual words are represented by using the distribution of the instances of medical visual words over the categories. Finally, medical images are represented by the vector sum of the instances. Cosine similarity is used to predict the category of an unseen medical image. The method combining with the visual word dictionary learning can simplify the construction of the class space, and can be easily extended to the hierarchical classification task. Both ImageCLEFMed database and Emphysema database are used in our experiment. The experimental results show that the performances of the medical image classification are improved by using the presented method. Especially the performance of the global medical image classification using hierarchical class space combining with the visual word dictionary learning is significantly improved.In conclusion, the proposed methods in this dissertation are effective to solve the issues in traditional medical image classification based on bag-of-visual-words model, and improve the accuracy, the sensitivity and specificity.