A robust segmentation method for counting bovine milk somatic cells in microscope slide images
Introduction
Somatic cells, such as leukocytes and epithelial cells, are always present in bovine milk. Inflamed mammary glands, also known as bovine mastitis, lead to considerable growth in the number of such cells. This biologic process causes economic losses to the dairy industry, once it affects the milk quality (Cavero et al., 2007). The National Mastitis Council (Bramely et al., 1996) estimated annual losses per cow in the U.S. of US$ 185.00 due to mastitis, and a total annual expenditure of US$ 1.8 billion. This is largely due to subclinical form of the disease (Wu et al., 2005).
Mastitis can be detected using various diagnostic techniques based on milk samples, and they are divided into direct and indirect methods. As examples of indirect methods one can cite the California Mastitis Test (CMT), the Wisconsin Mastitis Test (WMT) and the measurement of Electrical Conductivity (EC). Whereas examples of direct methods include counting of somatic cells by Direct Optical Microscopy (DMSCC) and electronic analyzers, such as flow cytometry or particle detection electronics (Moon et al., 2007). The Somatic Cell Count (SCC) is historically the most commonly used indicator for identifying mastitis. However, other indicators, such as electrical conductivity (EC), Neutrophil Count based on detection of O2, Lactate Dehydrogenase (LDH) and NAGase have been investigated in the last decades (Chagunda et al., 2006, Okada et al., 2009, Viguier et al., 2009, Ankinakatte et al., 2013). The EC of milk was introduced in 1940 as an indicator of mastitis (Davis, 1947, Norberg, 2005). Since then, several studies have been conducted to assess the accuracy of the EC and predict the state of infection (Hillerton and Semmens, 1999, Shoshani et al., 2000, Mele et al., 2001, Norberg et al., 2004, Norberg, 2005, Janzekovic et al., 2009, Ferrero et al., 2014).
Other studies have been conducted to develop indirect sensing systems for mastitis detection. For example, Eriksson et al. (2005) proposed a system based on gas-sensor and Wu et al. (2005) suggests another one using a sensor capable of detecting the Deoxyribonucleic Acid (DNA). Systems like these, together with artificial intelligence algorithms (AI), provide data that allow early detection of mastitis. For instance, Cavero et al., 2007, Miekley et al., 2012 implemented a detection system based on univariate indicator variables, whereas Kamphuis et al. (2010) applied decision trees to the problem. Cavero et al., 2008, Ankinakatte et al., 2013 have also developed systems based on neural networks, whereas De Mol and Woldt, 2001, Cavero et al., 2006, Kramer et al., 2009 devise a system based on fuzzy logic, and finally, Miekley et al. (2013) proposed another one based on PCA (Principal Component Analysis). However, beside all attempts to use AI and other algorithms, such systems based on detectors are vulnerable to noise, which has a considerable influence on the characteristics of the signal to be processed and the detection technique to be applied (Kamphuis et al., 2010). Therefore, there is still a strong need for improved detection methods that translate sensor data into reliable information to be processed.
Contrary to indirect methods, which provide only an indication of potential problems associated with high somatic cell count (Ferrero et al., 2014), direct methods are more accurate and recommended by the Food and Drug Administration (FDA). In the direct approaches, the levels of mastitis are determined by SCC per milliliter of milk (Janzekovic et al., 2009), and flow cytometers are one of the most widely used test systems for this purpose, due to the high throughput and accuracy. Their use has been standardized and recommended by international dairy organizations (Gonzalo et al., 2004, Garcia-Cordero et al., 2010). However, flow cytometers are expensive to purchase (>US$ 50,000), operate and maintain, precluding small farmers from adopting them for health routine monitoring of their herds. Typically, small farmers send milk samples to government agencies or centralized private laboratories to test them by SCC, along with other tests (Garcia-Cordero et al., 2010). On the other hand, it is possible to detect mastitis and the severity of it in milk samples by SCC performed via DMSCC. This is the reference methodology adopted by the International Dairy Federation and used even for electronic equipment calibration (Orlandini and van den Bijgaart, 2011).
Unfortunately, the manual DMSCC is a tedious task that requires slide preparation and two or more hours for visual counting by a human expert. Nevertheless, a considerable number of producers in many regions around the world cannot afford electronic counting equipment, and employing DMSCC is still necessary. Thus, the introduction of image processing techniques for automatic counting in DMSCC is of major importance for SCC.
Image processing is widely employed for recognition of different cells in the fields of biology and medicine, by using techniques or algorithms already standardized. There are several software available in the market that use image processing for counting of different kind of objects, particles and cells (Abràmoff et al., 2004). But, this kind of software has specific functions for very particular processing purposes, and each practical application requires the development of new algorithms or customizing the existing ones. For example, those software usually have specific functions for segmentation and counting of biologic cells in a general sense, but the expected limitations led Kong et al., 2011, Pan et al., 2012, Dorini et al., 2013 to develop new algorithms to deal with human blood cells. In the same way, Baro et al. (2005) presents a detailed description of the hardware and preparation methods for an automated DMSCC system (referred to as Video Microscopy – VM). This VM system is intended for SCC in cows’ milk, but no details of the employed algorithms are presented in Baro’s work. Moreover, recent works for somatic cell targeting have been proposed by Xue et al., 2008, Na and Heru, 2009, Xue et al., 2009, Na and Heru, 2010, Wang and Xue, 2010, but none of them has the necessary investigation for image variations.
The manual DMSCC technique is subjected to different manipulation procedures, depending on where it is practiced. Thus, the slide images obtained suffers many variations in their features, such as shade, hue, colorfulness, saturation, lightness and brightness, as well as the presence of debris. These variations are inherent to the process that precedes the capture of images and depend on the lactation period of the animal, hygiene conditions during the sample acquisition, staining and dyes types in the slide preparation, microscope illumination and slide region where images were taken from. Additionally, the variations in morphological patterns of somatic cells may deteriorate the performance of traditional segmentation techniques, since they can be efficient for a group of images but not for another one.
In this work, we propose a robust image processing algorithm for image segmentation to be used for SCC in automated VM system. The proposed technique is a combination of k-means clustering algorithm (Hartigan and Wong, 1979), Watershed transform (Roerdink and Meijster, 2000) and a new proposal for histogram thresholding. A set of one thousand DMSCC images was analyzed by a human expert and by the proposed algorithm. Promising results indicate that this technique may be used to obtain robust automated video SCC systems.
Section snippets
Slide Image Features
The slide preparation for microscopy of bovine milk consists of spreading 10 μL of milk in an area of 1 cm2, and left to dry out at room temperature to be colored with the Rosenfeld technique (Gondim et al., 1998). After completing this procedure, the slide is visualized and photographed in a microscope, using the proper magnification.
Fig. 1 exemplifies different aspects in the microcopy of somatic cells. As said before, cell morphology, debris and image variations are of major concern. According
Thresholding and clustering effects
k-means is an algorithm to group abstract objects with characteristics defined by n parameters, into p groups. Each group or cluster is obtained by minimizing the distance, in the parameter space, to the corresponding cluster centroid. This method has been applied to several problems of image segmentation, as presented by Date and Akarte (2013) for brain image segmentation. In the present application, the k-means algorithm is employed to cluster the original image into two patterns.
To
Conclusions
A new scheme for somatic cell counting in bovine milk has been proposed. Several cell images, captured from different slide preparation processes, was used to validate the experimental results. Based on the results of this study, it was observed that the use of a new thresholding, proposed for the use in unimodal histograms, produces better results than two traditional methods. The proposed scheme uses clustering, thresholding, segmentation and cell counting; providing very accurate results,
Acknowledgements
The authors would like to thank the Coordination for Improvement of Higher Education Personnel (CAPES) and the National Counsel of Technological and Scientific Development (CNPq) for supporting this work.
References (52)
- et al.
Predicting mastitis in dairy cows using neural networks and generalized additive models: a comparison
Comput. Electron. Agric.
(2013) - et al.
Mastitis detection in dairy cows by application of fuzzy logic
Livest. Sci.
(2006) - et al.
Analysing serial data for mastitis detection by means of local regression
Livest. Sci.
(2007) - et al.
Mastitis detection in dairy cows by application of neural networks
Livest. Sci.
(2008) - et al.
A linear-time component-labeling algorithm using contour tracing technique
Comput. Vis. Image Und.
(2004) - et al.
Detection of mastitic milk using a gas-sensor array system (electronic nose)
Int. Dairy J.
(2005) - et al.
Screening method for early detection of mastitis in cows
Measurement
(2014) - et al.
Canine hepatozoonosis in brazil: description of eight naturally occurring cases
Vet. Parasitol.
(1998) - et al.
Evaluation of rapid somatic cell counters under different analytical conditions in ovine milk
J. Dairy Sci.
(2004) - et al.
Comparison of treatment of mastitis by oxytocin or antibiotics following detection according to changes in milk electrical conductivity prior to visible signs
J. Dairy Sci.
(1999)
Decision-tree induction to detect clinical mastitis with automatic milking
Comput. Electron. Agric.
Mastitis and lameness detection in dairy cows by application of fuzzy logic
Livest. Sci.
Application of the ‘tracking signal’ method to the monitoring of udder health and oestrus in dairy cows
Livest. Product. Sci.
Detection of mastitis and lameness in dairy cows using wavelet analysis
Livest. Sci.
Application of fuzzy logic in automated cow status monitoring
J. Dairy Sci.
Application of a new portable microscopic somatic cell counter with disposable plastic chip for milk analysis
J. Dairy Sci.
Automatic thresholding for defect detection
Pattern Recogn. Lett.
Electrical conductivity of milk as a phenotypic and genetic indicator of bovine mastitis: a review
Livest. Product. Sci.
Electrical conductivity of milk: ability to predict mastitis status
J. Dairy Sci.
Leukocyte image segmentation using simulated visual attention
Expert. Syst. Appl.
Unimodal thresholding
Pattern Recogn.
A comparison of color models for color face segmentation
Proc. Technol.
Mastitis detection: current trends and future perspectives
Trends Biotechnol.
Deoxyribonucleic acid sensor for the detection of somatic cells in bovine milk
Biosyst. Eng.
Image processing with imagej
Bio. Int.
Video microscopy as an alternative method for somatic cell count in milk
J. Dairy Res.
Cited by (12)
A step forward in food science, technology and industry using artificial intelligence
2024, Trends in Food Science and TechnologyLymphoma images analysis using morphological and non-morphological descriptors for classification
2018, Computer Methods and Programs in BiomedicineCitation Excerpt :This step obtained a thresholded image as defined by f(x, y). Some of the cells that remained connected during the use of global thresholding can be separated by the use of the watershed transform with 8-connected neighborhood [25]. Therefore, we applied the watershed algorithm to the R channel image [26].
Somatic cell count in buffalo milk using fuzzy clustering and image processing techniques
2021, Journal of Dairy ResearchResearch advances in the automatic detection technology for mastitis of dairy cows
2023, Nongye Gongcheng Xuebao/Transactions of the Chinese Society of Agricultural EngineeringParticle Detection of Complex Images Based on Convolutional Neural Network
2022, Chinese Control Conference, CCCMobiloscope: A Technological Solution for Early Mastitis Detection in Dairy Cattle
2022, IEEE Latin America Transactions