Urine Sediment Image Segmentation And Recognition Of Two Kinds Of Particles

Automatic urine sediment classifier has significant influence on clinic urine analy-sis. As compared to the traditional manual way which is limited by the technical level, deviation error based on vision and low efficience, it can relieve the doctors of their hard, time consuming manual work and avoid diagnostic error caused by subjectivism. Moreover, it provides quantitative analysis and high efficience. Meanwhile, digital re-sults and pictures are convient for long-distance transfer which is important for long-distance medical treatment and consultation.Base on theories of digital image processing and pattern recogniton and some ex-periment, we develop a strategy for objects detection and recognition of two of them, which are cast and sperm. Our system is able to segment various sediments which are useful for diagnostic analysis from the backgroud of digital microscopic image. These sediments include white blood cell, red blood cell, epithclium, cast, sperm, crystal and mycete. The classifier of system can recognize cast and sperm from other objects.In image segmentation stage, first of all, based on the statistic information of stan-dard difference gradient image's histogram, we employ an bi-thresholding method. Second, a refining strategy adopting the combined canny and gray image information is selectivly applied to some lacal part of the previous binary image.At last, in terms of cell mass, in the light of priori knowledge of blood cells, we exlpore a method making use of distantce transform image information to locate each of the blood cells in the mass and thereby segment them from the mass.In feature selcetion, we extract some shape features according cast and sperm's physic characteristics. As for cast, we additionally design a texture feature.In classifier designing, classifiers based on dicision tree are developed for high ef-ficiency and simpleness. The classifier for cast combines shape features with texture features.A large number experimental results show that our strategy can extract objects from background neatly and precisely. Meanwhile, it can also achieve satisfactort recognition rates.