Semi-supervised Ultrasound Image Segmentation Based on Direction Energy and Texture Intensity

For the ultrasound images accurate segmentation problem, this paper proposes a novel SVM semi-supervised segmentation method based on major features in curvelet domain. Firstly, ultrasound images were decomposed into different directions and frequencies in the curvelet domain, then the cauchy model was used to simulate curvelet coefficients distribution, thus the main distribution of the curvelet coefficients were extracted to reduce the algorithm time complexity; Secondly after curvelet inverse transform we designed texture analysis method to distinguish texture intensity of every blocks among each sub-bands, then elected maximum K numbers energy sub-bands accroding to the strong texture characteristic, followly extracted features such as: angular second moment, contrast, correlation, entropy, variance, mean, adverse moments, etc from these maximum energy sub-bands, thereby calculating data amount was reduced and algorithm real-time performance was improved; Finally we designed semi-supervised SVM classifier and took the expert manual tagging map as reference standards, compared with the results of moment method and active contour model, experimental data show that our algorithm for ultrasound images pathological region segmentation has better accuracy and effectiveness.