Medical Image Retrieval Based On Local Texture Statistical Model

This paper aims at looking for an image retrieval method that can be effectively applied in medical images. After analyzing current image retrieval techniques that are widely used, I study on texture analysis methods based on Q-shift den droid complex wavelet transform and KLD similarity measure algorithm, starting from the texture analysis methods based on frequency domain. Then I focus on a medical image retrieval method based on local texture statistical model. The main contents are as follows:1. This paper analyzes and studies on the framework of medical image retrieval system and both the classification and characteristics of medical images. Some widely used image texture analytic procedures and similarity measurement algorithms are elaborated. Image texture analysis methods mainly include spectral analysis method, statistical analysis method and structural analysis method. Among them the former two methods are more widely used. Image similarity measurement algorithms include Euclidean distance method, histogram intersection method, L1-metric algorithm, Minkowski distance algorithm, Mahalanobis distance algorithm, etc.2. Starting from texture analysis methods based on frequency domain, This paper proposes a novel medical image retrieval algorithm based on the generalized Gaussian statistical model of Q-shift dendroid complex wavelet transformation. Compared with Flourier transformation, discrete wavelet transformation has some superiorities such as it can localize both time and frequency while multi-resolution analyzing signals. There exists, however, some disadvantages such as strong oscillatory behavior, frequency mixing, translation transformation inadaption and bad directional selectivity. To solve the above drawbacks, this paper brings in dendroid complex wavelet transform which has scale rotation invariance, translation invariance and no oscillation or mixing. In order to effectively reduce the computational load, here a medical image retrieval algorithm based on the generalized Gaussian statistical model of Q-shift dendroid complex wavelet transformation is proposed. The main idea of the algorithm is to describe the texture characteristics of medical images by scale and shape of which the distribution of the edge probablility of Q-shift dendroid complex wavelet transform sub-band wavelet coefficients approximately obeyes generalized Gaussian distribution model, then use KLD distance measure algorithm to measure the similarity and complete the retrieval. The simulation results shows that the precision obtained by this algorithm is 75.60%;3. In the foundation of statistical-based texture analysis methods, this paper studies on the local texture statistical model of images and proposes a medical image retrieval algorithm based on local texture statistical model of multifilter LBP pyramid. 9 classes, 17 classes and 25 classes uniform mode are simulated and analysed, respectively. It concludes that there are more equivalent modes contained in the 9 categories in the equivalent forms. The medical image retrieval algorithm based on local texture statistical model of multifilter LBP pyramid uses DOG filtering image to multiscale decompose images, then uses 9 categories of uniform LBP modes to establish the joint probability distribution model of LBP models and local gray variance to describe the image texture features, and finally adopts modified Log-likelihood statistics measurement to measure the similarity and complete the retrieval. In the simulation, the precision obtained by this algorithm was 99.08%. Since the algorithm only does texture analysis and similarity measurement for two images that are 1/4 and 1/16 size respectively of original image, the computation amount of the algorithm is greatly reduced than the general algorithms based on statistcal model and can be well applied in medical image retrieval.