Segmentation And Recognition Of Red Cells In Urinary Sediment Microscopic Images

Urinary sediment microscopy inspection is a common and important clinical test. Red cells in urinary sediment are critical subject of the test. Analysis of visible component in urinary sediment is currently carried out mainly by manual microscopy examination of the sample. This approach is laborious and subject to human impacts, and usually give qualitative results, thus is not able to fulfill the requirements for clinical quantitative test. With the development of computer vision and pattern recognition technologies, automatic quantitative analysis of visible components in urinary sediment becomes more and more feasible. The automatic analysis process can be parted into steps such as image preprocessing, segmentation, feature extraction and object recognition. The work presented in this thesis was mainly aimed at the segmentation and recognition of red blood cells in urinary sediment microscopic images.In the image preprocessing stage, the Mean Shift algorithm is adopted for color image smoothing. Canny edge detector is then applied on the smoothed image to obtain relatively intact object contours. Background noises in the image are also suppressed somehow.Usually, the red blood cells present in the images are in elliptical shape. Utilizing this fact, we adopt the well known parametric curve detection method, Hough transform, to find out these red cells. Since five parameters are required to describe an allipse, these too many parameters make the standard Hough transform unfeasible in our research. Rather, a randomized Hough transform is more suitable for our purpose. The sampling efficiency of the randomized Hough transform is analyzed, and the neccessity of reducing the noise level during the sampling is asserted. Based on the noise reducing idea, an improved randomized Hough transform for ellipse detection is proposed. A circular ring detection using Hough transform incorporating a priori knowledge on the red cell sizes is firstly performed in the whole image scope. This detection gives candidate positions where an elliptical red cell may exist. Randomized Hough transforms for ellipse detection are then carried out one by one on these candidate positions. To reduce the complexity of the algorithm with reasonable cost on memory usage, five 1D accumulator arrays are use to construct the Hough space of the Hough transform. Since in this case, the Hough transforms are restricted in relatively very small areas, the local signal-to-noise ratio is significantly increased, leading to the improvement in the efficiency and accuracy of the algorithm. Experimental results on real-world urinary sediment images show the effectiveness of the algorithm.Geometric features are extracted for the segmented elliptical object regions. A decision tree for red cell recognition is constructed by using the ID3 algorithm. The decision tree was applied on the real-world images to verification, and good results were achieved.