Detection, and counting of white blood cells in broiler chickens using image processing

Introduction: The analysis of blood cells is used for detecting several diseases in human and animals. White blood cells, also called leukocytes, are an important part of the body’s defense system against disease. In response to infection, the white blood cells (WBC) count typically increases (Sparavigna 2017). Therefore, measurement of the WBC count able to give an indication if there is infection or inflammation occurring. Increases in WBC can also occur in some cancer conditions such as leukemia. A bird under stress may also have a high level WBC, but greater elevations are a definite indication of disease. Reduced numbers of WBCs can occur due to bone marrow disease, severe acute disease and other conditions. White blood cell counts can be determined by various methods either through estimation or varied counting techniques. At this time, there is no automated system, such as is used with humans or dogs and cats that has proven effective for determining avian white blood cell counts (Sakas 2002); this is because of avian blood cells are nucleated, therefore, leukocytes cannot be separated from other blood cells. Bare nuclei from lysed cells, nucleated thrombocytes, and lymphocytes are all of similar sizes and difficult to distinguish by impedance or light scatter properties, and the contents of the leukocyte granules of birds differ from those of mammals. In addition, morphologic differences between avian species commonly limits the development of automatic protocols. These qualities limit the use of centrifugal hematologic analyzers such as quantitative buffy coat analysis, impedance, and laser cytometry tools to differentiate all types of cells (Harrison and Lightfoot 2005). Total WBC count is usually performed in laboratory by using manual techniques and hemocytometer. Differential counting has also been done by stained blood smear and light microscopy. Present of nucleated erythrocytes and thrombocytes in avian blood prevents from automated method that have been developed for mammals. Also, manual counting is time-consuming and requires to an experienced technician (Beaufrere and et al. 2013). The evaluation of blood smears for cell differentiation and detection of anomalies is a complex process and many morphological aspects of the cells are assessed and analyzed. Therefore, because human visual perception is a type of computation, it is possible that digital images can be analyzed by using a similar process. Content-based image indexing and retrieval has been an important research area in computer science for the last few decades. Many digital images are being captured and stored such as medical images, architectural, advertising, design and fashion images, etc. As a result large image databases are being created and being used in many applications (Maitra and et al. 2012). Therefore, this project carried out to examine the applicability of image processing system for rapid detection of white blood cell in broiler chickens. Material and methods: Blood samples were taken from 15 broiler chicks, and then blood smears stained with Giemsa color and finally were photographed by Canon E500. A total number of 80 RGB color image were created in jpg format, 42 images were selected and WBC counted by confirmation hematology expert. After that the images were processed by MATLAB R2017a and CellProfiler (CP 2.0 r10997) softwares. The appearance of WBC such as the color, shape and size were considered by MATLAB software. The images were examined in RGB, HSV and L*a*b color spaces. Visual examination and index recording showed that the color feature in all color spaces is not useful for detecting WBC. Since the cell nucleus was more distinct in a-layer images than other layers, the image of this layer was used for further processing. The a-layer images converted to binary images by a threshold. Then obtained binary image examined by the implementation of different conditional ring based on their size and shape. Two methods were used based on size; the first method is based on remaining four larger objects, and the second method removed the areas less than 5000 pixels in the image. In the shape method the interval 1-1.25 was considered to be form a factor for remaining the round objects and in the combined method, the removal of area less than 5000 pixels with an interval 1-1.25 have been considered. In the image processing by cellProfiler software, different modules and typical diameter 18-52 pixel for objects were used. Results and discussion: Among the different methods, the image processing with MATLAB software, and the use of second method (removed areas less than 5000 pixels in the image) had the highest number of diagnostic cells (84% of WBC were correctly detected). The rate of diagnostic cells of other method (first, third and fourth) were 35, 48 and 39% respectively. In the image processing by cellProfiler software, by using various modules and typical diameter 18-52 pixel for objects, WBC was identified in the rate of 84%. The results of our study show that our methodology was not consistent enough to agree with WBC counts obtained by using manual differential, which invalidates our initial hypothesis. Therefore, it is speculated that variability in the blood smear features, such as smear thickness, cell spread, and, therefore, surface area, cell touching, uneven spatial cell distribution (depending on areas of the blood smears), and the similar appearance of thrombocytes and small avian lymphocytes could contribute to the results. Color error and nucleated red blood cells are the biggest problem in this project and more advance algorithms are required to solve this problem. Based on our results it seems that nothing can replace the expertise of a trained technician or veterinarian, but manual CBC determination is time consuming, partially subjective, and semi quantitative. Despite tremendous advances in the machine learning, recent investigations have shown that humans have better visual categorization abilities than machines. However, machines are better, for instance, at computing complex geometric properties, performing more precise measurements and statistical evaluation, and evaluating complex sets of rules, and they can adapt quickly from training samples. Conclusion: Although our technique did not provide satisfactory results in its present form, we hope that this approach can still stimulate further research in the application of digital image analysis. Generally, the results showed that more advanced machine-learning algorithms and neural networks are required to be test for counting and differential white blood cell diagnostics.