Recognition of Marrow Cell Images Based on Fuzzy Clustering

Published Online February 2012 in MECS DOI: 10.5815/ijitcs.2012.01.06

• 1 ．    Introduction

The counts of different types of white blood cells in bone marrow, the so-called differential counts, provide invaluable information to doctors in diagnosis of diseases such as AIDS, leukemia or cancers. The traditional

Shanghai International Science and Technology Cooperation

Foundation Project (11140903700).National Nature Science

Foundation of China (81170507). Corresponding Author: SHI Jun,

method for an expert to achieve the differential counting is tedious, error-prone and time consuming. Analyzing the blood cell images is also important in clinics. This is especially true for the leukemia which is known as the most dangerous disease. We believe that the shape and the number of leukocytes in the bone marrow cell image are the important information for the diagnosis of leukemia. Our previous works are focused on the human marrow cell image analysis [3, 4], which can roughly get the count of each kind of cells. However, the experimental data is based on the patients’ test data, and the progress of leukocyte is usually unknown before we do our visual identification. In order to keep a close tracking of the development of leukemia, we use a group of mouse to watch closely on the leukocytes developments. Methods similar to those of human beings are used to identify the different kinds of objects on the marrow sliced plates. As the staining difference and color difference, the different threshold values are used to retrieve the configuration of the cells. Similar fuzzy theory and iteration steps are followed for the whole processing. It is expected that an automatic discriminating system can be set up to save time and will let the investigation more efficient.

We expect the number of myeloblast and promyelocyte will out match the number of metamyelocyte. We also expect that the distance between these cells will be different to those later stages from earlier healthy stages. So it is important to get the cells counts of the different cells in the marrow samples. While most of these kinds of identification can be done by cell staining and then the specific color picking, our experiment has met a problem that the color can not be efficiently attached to those interested cells (or neucli), only one or two colors are viewable in the micro-scope image. Before we present our solution to this problem, we will cover some of the similar works done on good stained samples.

Many methods for recognition and segmentation of blood cell images have been proposed. Most of them utilized the gray level, texture, and color. In reference [1], a new, fully automated, content-based system is proposed for knee bone segmentation from magnetic resonance images (MRI). The purpose of the bone segmentation is to support the discovery and characterization of imaging biomarkers for the incidence and progression of osteoarthritis, a debilitating joint disease, which affects a large portion of the aging population. The segmentation algorithm includes a novel content-based, two-pass disjoint block discovery mechanism, which is designed to support automation, segmentation initialization, and post-processing. The block discovery is achieved by classifying the image content to bone and background blocks according to their similarity to the categories in the training data collected from typical bone structures. The classified blocks are then used to design an efficient graph-cut based segmentation algorithm. Content-based refinements and morphological operations are then applied to obtain the final segmentation. The technique does not require any user interaction and can distinguish between bone and highly similar adjacent structures, such as fat tissues with high accuracy. The performance of the proposed system is evaluated by testing it on 376 MR images from the Osteoarthritis Initiative (OAI) database. The results show an automatic bone detection rate of 0.99 and an average segmentation accuracy of 0.95 using the Dice similarity index.

In reference [2], the active contours without edges model is applied to compute the segmentation of an image into two phases. The minimization problem is non-convex even when the optimal region constants are known. The paper applies a method that can compute global minimizes by showing that solutions could be obtained from a convex relaxation. A convex relaxation approach is further proposed to solve the case in which both the segmentation and the optimal constants are unknown for two phases and multiple phases. So a relaxed convex of the popular K-means algorithm is used which can compute tight approximations of the optimal solutions.

In reference [3,4], probability and fuzzy set methods are used to retrieve cell features from image, and process automatic counting of various marrow cells that are in-sufficiently stained. Samples are based on checking records of a series of lukemia patients. The problem is that when the image quality is good, the recognization rate is good; otherwise, the algorithm will fail. To resolve this problem, higher resolution lenses are used and larger sizes of pictures are used, the expectation is that the combined image will be more efficient when the similar searching and matching algorithms are applied.

In reference [5], A SIFT algorithm in spherical coordinates for omnidirectional images is proposed. The algorithm can generate two types of local descriptors, Local Spherical Descriptors and Local Planar Descriptors. With the first ones, point matching between two omnidirectional images can be performed, and with the second ones, the same matching process can be done but between omnidirectional and planar images. Furthermore, a planar to spherical mapping is introduced and an algorithm for its estimation is given. This mapping allows to extract objects from an omnidirectional image given their SIFT descriptors in a planar image. This kind of method is useful when the current project advances to the stage that 3D data need to be processed. This is important to resolve the problem that the slicing is random and a lot of useful information can be miss-reading because of the incomplete cell parts.

In reference [6], a recognition method for the blood cell images was proposed. Since red cells, leukocytes, platelets, and cytoplasm had different color in the blood cell image, they were extracted according to their own colors. First, the color areas of red cells, leukocytes, platelets and cytoplasm were determined, respectively. Second, pixels were distributed into each color area by using the fuzzy clustering algorithm. The leukocytes, platelets, and red cells were detected accurately in all five images.

In reference [7], a method based on the color fuzzy clustering was proposed to divide the color areas and distribute each pixel into each area. The technique will segment single cell images of white blood cells in bone marrow into two regions, i.e., nucleus and non-nucleus. The segmentation is based on the fuzzy C-means clustering and mathematical morphology. The segmentation results are compared to an expert’s manually segmented images. The initial investigation of the use of the derived segmented images in the cell classification is also performed by using the Bayes classifier.