Segmentation And Recognition Of Casts In Urinary Sediment Microscopic Images

Urine analysis is one of the common clinical tests. Generally, the urinary sediment analysis is supposed to give quantitative results on visible components in the urine sample, such as red blood cells, white blood cells, calcium oxalate crystals, casts, fat granules, bacteria, etc. These results are supportive for diagnosis of the urinary system and the circulatory system diseases or damages. Cast is an important type of the urinary visible components, whose existence can be used to determine the health of the kidney, and is thus of important implications for clinical diagnosis.Traditional urinary tests are carried out through manual microscopy. This approach is laborious, introduces subject human impacts into the test, and is hard to obtain quantitative results. During the manual process, the expertise, mental situation, and the physical weariness of the doctor may cause inaccuracies in the test results. The precision and the reliability of the manual test is thus reduced.With the development of computer technologies, considerable progresses have been witnessed in the computerized image processing technologies. Analyze urinary sediment microscopic images by image processing and machine vision techniques, the casts in the urinary sediment can be quantified automatically. This can lead to the improvement in the accuracy of the clinical test, and can release the medical stuffs from the work burden. Researches on cast segmentation and recognition based on image processing is thus of applicable importance. Besides, the relevant segmentation methods can be helpful for other image types, and is then nontrivial in its theoretical implications.The work presented in the paper utilizes digital image processing and pattern recognition techniques to segment and recognize the casts in urinary sediment microscopic images. The main contributions of the work is as follows.(1) A cast segmentation method based on Mean Shift and the improved maximum entropy segmentation is proposed. Images are smoothed by using the Mean Shift algorithm, and histogram equalization is performed on the smoothed images for enhancement. These steps result in images that are suitable for the afterward maximum entropy segmentation, and better segmentation can be achieved. Once the coarse segmentation is done by using the maximum entropy segmentation method, the holes in the obtained regions are filled out, and the binary regions are treated by mathematical morphological operations, which give the candidate cast regions for later recognitions. Compared with some of the existing methods, the proposed method can give better segmentation of the casts in the images.(2) Geometric features of the candidate cast regions, such as the longest axis length, the shortest axis length, the perimeter and area, are extracted by using the basic circumscribed rectangle method. The ID3 algorithm is then employed to construct a decision tree for recognizing casts. Good recognition results were achieved on real-world urinary sediment microscopic images.