Liver ultrasound is a cost-effective, non-invasive, and sufficient technique to diagnose most of the liver disorders. The recent advancements in research in image processing have led to the development of image-based liver disorder classification systems. In spite of being popular in the diagnostic imaging of liver, ultrasound images, owing to their poor quality, render the conventional and state of the art segmentation and feature extraction techniques incapable, to accurately classify a large mixed group of liver disorders; due to the similarities and differences in appearances among the different and same disorders, respectively. Classification of liver disorders using ultrasound images poses various challenges at each phase, from segmentation to classification. There is a need for better segmentation, powerful features, and optimal classification parameter combinations to obtain decent classification accuracy, when a large sub-set of liver disorders is considered. In this work, the region of interest is extracted using iso-contour technique. Feature extraction is performed using multi-level fractal features and multi-domain wavelet-texture features for better discrimination capability. Then, an optimization problem is formulated, for minimizing the five fold cross validation error to classify 10 types of disorders, both focal and diffused, by selecting the best features, suitable classifier, and its parameters using the particle swarm optimization technique for obtaining better classification. An overall accuracy of 91% is obtained using the proposed features in addition to 50% reduction in multi-level fractal feature set which justifies the efficacy of the proposed technique.

中文翻译：

---

基于粒子群优化的肝脏疾病超声图像多级多域特征分类

肝脏超声是一种经济高效、无创且足以诊断大多数肝脏疾病的技术。图像处理研究的最新进展导致了基于图像的肝脏疾病分类系统的发展。尽管在肝脏诊断成像中很受欢迎，但超声图像由于质量差，使得传统和最先进的分割和特征提取技术无法准确分类大量混合的肝脏疾病；由于不同和相同疾病分别在外观上的异同。使用超声图像对肝脏疾病进行分类在从分割到分类的每个阶段都面临着各种挑战。需要更好的分割，强大的功能，当考虑大量肝脏疾病子集时，最佳分类参数组合以获得不错的分类准确度。在这项工作中，感兴趣的区域是使用 iso-contour 技术提取的。使用多级分形特征和多域小波纹理特征进行特征提取，以提高辨别能力。然后，制定了一个优化问题，通过使用粒子群优化技术选择最佳特征、合适的分类器及其参数来最小化五倍交叉验证误差，以对 10 种类型的疾病进行分类，包括局灶性和扩散性，以获得更好的分类.