The Research On Techniques Of Feature Extraction For Hepatic CT Images And Its Application In Retrieval

With the use of multimedia technology widely in the medical field, various medical images processing systems are playing an important role in the clinical, teaching, research, medical image storage, retrieval and communication systems. The function of extensions such as image retrieval, auxiliary diagnosis in PACS is the necessity of development of PACS. Traditional medical image retrieval is text-based query to retrieve medical images by matching exactly database fields or image annotations. Text-based queries face the difficulties of manual annotation for medical imags. Due to several disadvantages such as unrepeatability, being sensitive tosubjective influence, generating heavy workload of manual annotation, text-based methods have been unable to meet the nees of large-scale medical image database retrieval. Thus the techniques of content-based image retrieval (CBIR) for medical image database queries have been major topics instead of text-based searching techniques in recent years.Because the CBIR technology that only makes users to choose the interested visual images research all the same and similar images in a large medical image database, so it have a much-needed to access image database through visual content in teaching and studying. As we see, the visual features in the medical examination are another interesting focus in the field of medical research. CBIR with visual features not only query the same or similar medical diagnosis, and also query the cases that are similar visually and different in diagnoses result. In teaching, CBIR technology can help teachers and students to browse the image database, and identify query results visually.However, medical images hasve some disadvantages of the low resolution, high-noise, only gray signal, automation processing hard, which makes CBIR for medical images still face with great challenges. In the process of diagnosis for radiology images, clinic decisions are usually based on region of interest (ROI). Moreover, the existing medical image processing technologies are not mature enough to make satisfing diagnostic results. How to combine the key technologies in image processing with medical images become the main research objective for physicians to provide scientific, convenient and accurate medical instruments.All the works in the thesis are carried out under Major State Basic Research Development Program (No:2010CB732500) and the key projects supported by national natural science foundation of China (No:30730036). They focus on the study of feature extraction technology of medical images, which include mainly the globle and local (ROI) feature extraction algorithms and image retrieval based on these algorithms.The main contributions of this dissertation are summarized as follows:1. A global feature extraction algorithm of hepatic CT images is proposed based on the non-tensor product wavelet, which can be used to the computer-aided diagnosis (CAD) in medical CT images with lesion. With the lackness of direction after separable wavelet transform, four-channal non-separable two dimension wavelet filter banks are constructed, and used to decomposition in hepatic images. The histogram of wavelet coefficients in each high-frequency subband are approximated using generalized Gaussian model (GGM). Experimental results show that this method can improve the retrieval results of lesions.2. A global feature extraction algorithm of hepatic CT image is proposed, which aims to improve the detection rate of liver lesions. The proposed method uses the low frequency subband coefficient of wavelet decomposition based on the non-tensor product coefficient. These coefficients are modeled by Gaussian distribution piecewise. which utilizes the approximate characteristic of low frequency subband and could better express the global feature of hepatic image. The experiment on1688CT images confirms that retrieval perform based on this algorithm have robust result.3. With respect to lesion of liver (ROI), two kinds of local feature extraction algorithm are proposes considering the behavior of liver cancer, liver hemangioma, liver cyst images after enhancement, that is, enhancement of liver cancer can described as "fast in and fast out", and enhancement region of hemangioma gradually filled to the lesion center, and cysts have no change in lesion. First we introduced the role of gradient image to extract feature points in four directions of the image gradient values. Second, distance transform is carried out on ROI, and the lesion is the fractal dimension features are then compute from the image data of every layer due to the characteristics of self-similarity of fractal dimension. Experiments show that retrieval result based on these two algorithmshave greatly improved. To further improve the retrieval performance, measure of similarity distance learning were studied in this paper. We mainly study Mahanalobis distance measure learning algorithms about k-nearest neighbor classification.The algorithms about Boostmetric is introdcuced detailed, and three learning algorithms LFDA, LMNN, BOOSTMETRIC are used in the liver tumor image.4. Two facts were summarized taking account of the characteristics of triple-phase contrast-enhanced CT images. First, the lesions of HCC and hepatic hemangiomas mostly had specific characteristic changes, whereas no change occurred in that of cysts. Second, the surrounding liver parenchyma information of lesion was considered because of the discrepancy in density between the lesion and the adjacent normal parenchyma in triple phase scans. Thus, according to these analyses and the studies in the previous chapter, a feature extraction algorithm of lesions came to being considering the specific behavior of focal liver lesions and their surrounding liver parenchyma after enhancement. The proposed feature algorithm originated from distinct imaging characteristics of lesion images and the surrounding liver parenchyma in triple-phase CT images, which was also the diagnosis perspective of clinicians or radiologists for three types of tumor patients. The algorithm mainly included two processes. The first was distance transformation, which was used to partition the lesion into distinct regions and this represented the spatial structure distribution. The second was bag of visual words (BoW) representation based on regions. The latter was the key step. Generally, the lesion was divided into three regions in our experiments, which will fit best with imaging analysis of radiologists in triple phases for the diseases described above. The effect of the number selection on the performance of CBIR will be discussed in this article. The study’s aims are to develop a feature extraction algorithm of hepatic lesions considering the radiologists’ diagnosis views in triple-phase enhanced CT, and to contribute to a CBIR system to facilitate the retrieval of radiologic patients whose lesion images had similar-appearing with the query patient, and to implement a basis evaluation of this system.Research around feature extraction of liver CT images is discussed in this dissertation, mainly in study of the globle and local (ROI) feature extraction algorithms and image retrieval based on these algorithms. Finally, a content-based medical imge retrieval system for liver CT images is designed, and the results indicate that these methodes can improve retrieval accuracy.