Study On Several Problems Of Ultrasound Medical Image Processing

ABSTRACT Left ventricular (LV) volume measurement is a typical application of ultrasoundmedical image segmentation. Blurred edges in echocardiograms make themeasurement very difficult. Aiming at this application, firstly, noise suppression andedge detection methods were studied based on wavelet transform, and then, texturesegmentation methods were developed on this basis. On the basis of assimilating the former achievements, we applied severaltheories and methods, such as wavelet transform, fuzzy sets, rough set, Markovrandom field, feature space clustering, in ultrasound image processing. The maininnovations and improvements we have made in this study are as follows: Firstly, the wavelet shrinkage technology was improved, in which the shrinkagewas not processed in the domain of wavelet coefficients but processed in the domainof wavelet vectors'modulus. In addition, the relationship among the thresholds ofvarious scales was derived, and a uniform threshold formula was presented. Secondly, rough set theory and rotating neighborhood technology were used inthe noise suppression. Noise rough set was defined according to condition attributewe defined. The upper approximation and the lower approximation set wereseparately used to suppress noise. An isolated point noise suppression method, anedge detection method, and a region edge detection method based on rough set androtating neighborhood technology were developed and applied in the texturesegmentation after processing. Thirdly, a wavelet fuzzy operator edge detection method was proposed. In thismethod, wavelet multiscale edge detection and fuzzy edge detection were combined,multiscale edges, being the results of wavelet multiscale edge detection, became theobject of fuzzy processing. A fuzzy membership matrix (fuzzy plane) was constructedby defining a discrete membership function to image edge set. A fuzzy enhancementoperator with changeable crossover point was applied in the enhancement of thefuzzy plane to get a certain image edge set. This method took the advantages ofmultiscale of wavelet transform and ability to cope with uncertain problems of fuzzysets; as a result, it resolved the conflict between localization and detection in edgedetection. Fourthly, a gradient wavelet texture model was presented based on Gauss IIIABSTRACTMarkov random field and gradient wavelet transform. The parameters of the modelforms a multiscale texture feature space, and provided a multiscale description ofimage texture. K-means algorithm was applied in the clustering of the parameters,and in the process of clustering the result of coarser scales acted as the initialclustering center of the finer scales, as a result, the efficiency of clustering wasimproved. This method took the advantages of Markov random field image modeland imported the multiscale feature of wavelet transform. In this dissertation objective evaluation methods were adopted. The result ofevaluationindicated that the wavelet shrinkage and noise suppressionbased on roughset could suppress speckle noise in ultrasound images effectively, and the waveletfuzzy operator edge detection method had a good detection and noise-attenuatingability, and the texture segmentation based on the gradient wavelet texture modelcould give an exact segmentation of areas with different texture. In addition, wefound that edge detection methods were accomplished in the processing of imageswith complex structure and light noise, on the contrary, texture segmentation methodswere accomplished in the processing of images with simple structure and severenoise.