Research On Segmentation And Feature Extraction In Urinary Sediment Microscopic Images

Urinary sediment microscopy analysis is one of the methods for human medical examination in clinic. In recent years, urinary sediment auto-analysis methods based on image processing techniques have become a trend in biological cell recognition, which is important on theory and application. However, because of the imaging system defects and the specific mechanisms, the segmentation of urinary sediment images is a difficult job when come to the weak edge particles, in addition, the feature extraction is also a challenging work due to the multifarious species. In this paper, segmentation and feature extraction methods in urinary sediment images are studied and the main research works are as follow:(1) To deal with the weak edge images, a developed local rank transformation(DLRT) method is presented. In this method, a new judgment in local rank transformation is introduced by using the standard deviation in the neighbourhood, and the product of the standard deviation and the local rank is serve as the transformed value for the current pixel. A local rank transformation map will be formed through an iterator and the binary image is obtained through a thresholding algorithm. Then, a fine segmentation step is needed for the non-weak edge particles by utilizing Canny algorithm and morphology operation. Experimental results on different types of weak edge images show that the proposed method achieves high robustness in image segmentation.(2) To address with the feature extraction problem, a local histogram statistics(LHS) method and a distance gray level and differential cooccurrence matrix(D-GDCM) method are proposed, and while integrating with some geometric feature, the feature vector is created. The LHS and D-GDCM methods are presented as follow:1) In order to describe the particle characteristics that are invariant to rotation and illumination variation, a LBP-based local histogram statistics(LHS) method is presented. First, a rotation invariance with uniform LBP descriptor is introduced and yield the LBP map of the raw urinary sediment image. Then the raw image and map is sliced to three parts by using distance transform. At last, histogram of all the parts are calculated to achieve the LHS descriptor.2) In order to improve characteristics' abilities and raise the robust of our recognition system, a distance gray level and differential cooccurrence matrix(D-GDCM) method based on second order statistics analysis is proposed. First, the image center is determined by calculating moments of the image and so as to obtain the distance map. Second, the map and the raw urinary sediment image are jointed to collect statistics, which generates the distance gray level cooccurrence matrix and yield the feature vectors. Four gray-value differential invariants that is the output of “local jet” operator of the raw image are also take into consideration, to generate the distance differential component cooccurrence matrix in the same way. Finally, all the vectors are cascaded to produce the D-GDCM feature.(3) The identification of seven kinds of urinary sediment images such as epithelial cells, casts, leukocytes, erythrocytes, crystals, yeasts, impurities is achieved by using the tool of SVM. The experiment results demonstrate that LHS and D-GDCM are equipped with powerful discrimination capability that are invariant to rotation and illumination variation, and are very suitable for the identification of urinary sediment images under complicated conditions. 91.3% of precision rate and 91.5% of recall rate are achieved by using the proposed method.