Integration Of Global And Local Features Of The Medical Image Classification

With the development of medical imaging technology, more abundant and detailed information is provided by medical imaging for modern medicine, and various modes of medical images (such as, X-ray, CT, MRI, etc.) are playing an extremely important role in clinical diagnosis, teaching and scientific research. In recent years, digital medical images are applied widely, and the number of digital medical images is also increasing rapidly and accelerated. How to find the needed images from a large volume of medical images data is becoming an urgent issue. However, the medical image retrieval technology is unsatisfied, and becomes a bottleneck for full use of digital medical images. For a heterogeneous medical image database, if the medical images are classified by the image modality, body parts, organs and other attributes, it can effectively improve the image retrieval performance. Although the DICOM header contains some attribute information, DICOM header information has a high error rates, recent studies indicate that the attribute, body part examined, in DICOM header has an error rate of 16%. Therefore, automatic medical image classification is being recognized as an indispensable part for large medical image retrieval system. It can narrow the semantics gap, filter out non-relevant categories in the retrieval process, and reduce the search space, accordingly to enhance the retrieval performance.Automatic medical image classification is to give the semantic category labels to medical images, which can be regarded as a supervised learning process. It learns the mapping from image features to semantic categories by some machine learning algorithms to classify corresponding images to pre-defined categories. Automatic medical images classification contains two parts: image feature extraction and classifier construction. Most image classification algorithms tend to use either global features or local features to represent images. Global feature is calculated from all pixels of an image and describes the image as a whole; local features are used to describe the local details of an image, and always more robust to lighting and occlusion. The two types of features provide different information about the images, so it is reasonable to combine global feature and local feature to improve the classification accuracy. Based on the analysis and comparison of a variety of global features and local features, this dissertation explores and studies two medical image classification algorithms, which combining global features and local features. They are named as low-level feature fusion and high-level feature fusion. Low-level feature fusion concatenates different features to form a feature vector; High-level feature fusion fuses the classification results based on different features. Experimental results show that our methods can effectively improve the accuracy of medical image classification. The main contributions of this dissertation are:(1) The state of the art in medical image classification is reviewed from the two perspectives, feature representation and classifier, and the trends of this research area are exhibited.(2) Virous common used global features are analyzed, and their performances in medical image classification are compared by experiments.(3) Current popular local features are analyzed, and their performances in medical image classification are compared by experiments.(4) Several superior global features and local features are choosed based on the previous comparison studies, and two feature fusion methods are explored. The experimental results show that combining global and local feature can boost medical image retrieval, and combining LBP and ModSIFT features with high-level feature fusion can achive the best performance.