Chest X-ray Retrieval Based On Texture Feature Description And Matching

With the rapid development of medical imaging equipment and computer technology, the number of clinical images acquired all over the world increases explosively. To manage and transit clinical images, Picture Archiving and Communication Systems (PACS) are established in most middle-upper level hospitals for physicians. In addition, the modality diversity and complexity in medical images give rise to a requirement to develop a talent way to assist doctors in diagnosis as well as medical education programs. Retrieving medical images from an image database archived in PACS is one solution to aid radiologists and physicians. In recent years as the content-based image retrieval (CBIR) is developed, the relevant research in medical field, called content-based medical image retrieval (CBMIR), has reached a few research-oriented achievements and has been used in some research institutes. CBMIR contains two main parts:visual feature description and similarity matching.In visual feature description, this work discussed several feature extraction methods which represent image texture, according to the characteristic of chest X-ray images. Texture is a feature related to the pixel distribution spatially, which is more suitable to human visual system compared to gray scale and shape. In this work, texture feature extraction with the Gray-level Co-occurrence Matrix and the multi-scale Gabor filter bank was introduced respectively, together with the Scale Invariance Feature Transformation (SIFT) algorithm. The retrieval accuracies by these feature extraction methods were compared.In similarity matching, different techniques for feature matching were applied, according to different feature extraction methods. The global image feature vectors generated by Gray-level Co-occurrence Matrix and the multi-scale Gabor filter bank are matched by Euclid Distance, and a self-adapted weighing distance metrics method was raised. The key points and there feature vectors extracted by SIFT are matched by a Kd tree searching and nearest neighbor matching method.The radiograph retrieval using feature extraction and similarity matching methods are applied and realized. The results indicate that compared to the typical Gray-level Co-occurrence Matrix, the feature extracted from responses by multi-scale Gabor filter bank perform better; the self-adapted weighing distance metrics method raises the retrieval accuracy; and feature vectors extracted by SIFT are matched by a Kd tree searching and nearest neighbor matching method have a significant retrieval accuracy dealing with radiographs with larger lesions.