Reasearch On Feature Combination And Feature Learning For Medical Image Classification

Along with the popularization of advanced medical equipment, medical images have increased exponentially, and because of tons of image data, there are both opportunities and difficulties in medical domain. On the one hand, the original management techniques have been antiquated due to the production of image data, it is do essential to handle the medical images with a more effective method, on the other hand, as one of the challenges in medicine, the prevention of nodule proliferation and raise of survival in lung cancer patients make early diagnosis reasonably significant. However, there is a great deal of difficulty for therapist to identify the tumors of which feature is obscure in lung. Fortunately, with the exploitation of pattern recognition and machine learning, we could construct a CADs (Computer-aided Diagnose system) used to locate the nodule in X-ray images.To solve two different aforementioned problems on the classification of medical image, this thesis is focused on two aspects, one is the research of medical image modality classification with combining visual and text features, the other is the research of detection and identification of lung nodule automatically in chest X-ray image with feature learning. Firstly, this study reviews the research status of several crucial algorithms related to two research points, then combines visual and text feature of medical image with Lp-norm multiple kernel learning with respect to the first research point, and furthermore we achieve it with the best result based on IMAGECLEF2010 dataset. Concerning the second research point, we complete the detection of candidate nodules by image processing methods, followed by training the classifier with Single-Layer networks, and then implement the classification of nodules in chest X-ray images successfully. Here are four aspects about the main contribution of the paper:(1) There arise the curse of dimensionality problem when extracting the text feature of medical image, and the paper conducts the feature selection based on IG (Information Gain) measurement for text feature dimension reduction.(2) To achieve the medical image modality classification, the paper extracts four visual features and text feature and combines them with Lp-norm multiple kernel learning.(3) Due to the large amounts of candidate nodules, the paper constructs two classifiers, one model is used to the preliminary selection of candidates, and the other is trained with samples which are classified in error, and finalizes the property of nodules.(4) To terminate the recognition of lung nodules with a good result, the paper trains the final classifier with feature which obtained by applying to Single-Layer unsupervised feature learning algorithms, including Sparse Auto-encoder, RBMs and K-means.