Automatic Recognition Of Particles In The Urinary Sediment

The particles in the urinary sediment represent various and complex characteristics, it is difficult to acquire the features rapidly and accurately which can represent the various shapes and textures due to the problems such as unclear focus of the urinary sediment image, uneven illumination, blur, background complexity and other issues, which are likely to cause the image of the urinary sediment particles showing different features. Many atypical physical components in urinary sediment such as diverse impurities and cell debris, if not addressed, will often disrupt the recognition of the components of the normal elements. These problems have increased the difficulty of automatic recognition of the urinary sediment and affected the accuracy. In this paper, the author proposes several effective recognition algorithms by using digital image processing and pattern recognition techniques to recognize the particles in the urinary sediment, both in theory study and practice.At first, in order to extract features better, the Canny operator is adopted to detect the edges in the urinary sediment images, and then the binary images are obtained by combining with the morphological operations. On account of that the image blurring and uneven illumination would lead to incomplete Canny edges, an algorithm is proposed to improve the performance by calculating the standard deviation of the image gradient and then dynamically adjusting the threshold of Canny operator based on the result. Thereby, relative complete binary images can be obtained through this method, even when the images are of different quality.In the aspect of feature extraction, both texture features and shape features are extracted. Shape features include the area, the perimeter, the circular degree, the aspect ratio of minimum bounding rectangle, and so on. Especially on account of the distinctiveness of yeasts sprouting to balls, we put forward a method as follow: firstly, we detect the concave of yeasts so as to separate every single ones, and secondly we describe their shape features by incorporating with ellipse fitting. In terms of texture feature, a novel Local Jet texture feature based on distance transformation is represented. Firstly, we describe the local texture in different multi-scales by third-order differential of the Local Jet, and then count these texture features hierarchically in the light of the range of distance. The extracted texture features are invariant with respect to rotation, shift, and certain scale, and can even avoid the interference of background. These features also show the local texture distributions in the image, so they are effective to represent the texture of the particles. The experimental results demonstrate that the features can recognize various kinds of the particles in the urinary sediment images effectively, particularly to impurities of atypical particles, so that the interference to the recognition of typical particles can be excluded.In the aspect of classifiers, the Support Vector Machine (SVM) is applied in the texture feature classification. According to the characteristics of different particles in the unitary sediment images, we utilize decision tree to combine the texture and shape features together, and recognize 6 categories of particles effectively, including red blood cells, white blood cells, epithelial cells, mucus, yeasts and impurities.