Image Analysis Approach for Development of a Decision Support System for Detection of Malaria Parasites in Thin Blood Smear Images

Introduction

Malaria is a serious disease prevalent in developing countries. The World Health Organization estimates 300–500 million malaria cases and a few millions of deaths per year [1, 2]. It is a leading cause of mortality and morbidity in tropical and subtropical countries. Hence, it is an important health problem which needs to be addressed.

Malaria is caused by protozoan parasites of the Plasmodium genus by entering the bloodstream. The four species of malaria parasites infecting humans are Plasmodium falciparum, P. vivax, P. ovale and P. malariae. P. falciparum, which is found worldwide in tropical and subtropical areas, is the only species that can cause severe, potentially fatal malaria. P. vivax is found mostly in Asia, Latin America and in some parts of Africa. Because of the population densities especially in Asia, it is probably the most prevalent human malaria parasite. P. ovale and P. malariae are less frequently encountered.

Many techniques have been developed to diagnose malaria such as flow cytometry, fluorescent microscopy, PCR, etc, but light microscopy remains the gold standard for laboratory confirmation of malaria. Malaria parasites can be identified by examining a drop of the patient’s blood under the microscope, spread out as a “blood smear” on a microscope slide. Thick and thin smears of Giemsa-stained peripheral blood samples are used for diagnosis. Thin smears are used to assess degree of infection and species of malaria parasite. The specimen is stained with Giemsa stain, which expresses the parasites in a distinct manner. Giemsa stain is used to differentiate nuclear and cytoplasmic morphology of platelets, red blood cells, white blood cells and parasites. Giemsa staining solution stains up nucleic acids, and therefore those components containing the nucleic acids, namely parasites, white blood cells and platelets, are highlighted in a bluish purple color. Red blood cells are usually colored in slight pink. The rings of the trophozoites take up pale blue color.

The parasitemia is measured by counting the number of parasites present in the specimen. The hematologist has to study about 100 to 300 microscopic views of the blood smear to detect the parasites and evaluate the extent of infection. This is a subjective, laborious and time consuming task. It has been reported in [3] that the agreement rates among clinical experts for the diagnosis are surprisingly low. Moreover, the accuracy of diagnosis depends on quality of reagents, microscope and on experience of the laboratorian. Malaria is a seasonal and contagious disease, and hence it needs to be brought under control as early as possible. This calls for rapid and accurate diagnosis. Also, handling huge number of cases in a short time is a challenging task. In many developing countries, microscopy is not reliable since the microscopists are not well trained and are overworked. Conversely, in endemic countries, laboratory technicians are often unfamiliar with malaria and are likely to judge the parasitemia inaccurately. Thus, it is important to automate the diagnosis of malaria parasite detection so that the burst and spread of malaria can be controlled. Automation could play a very important role in the case of mass screening for malaria. It is very important to produce a common standard tool which makes the diagnosis based on a certain criteria and thus eliminates subjective assessment and makes the procedure highly repeatable.

The recent convergence of technologies is opening up new avenues for telemedicine applications. In many parts of the world, more so in developing countries, the rural areas without sufficient healthcare infrastructure or trained professionals are often well covered by mobile phone service providers. There is a lot of scope to develop innovative low cost technologies which can have a good impact on the healthcare systems of the developing countries. There have been efforts to use mobile phone based diagnostic systems [4, 5] for diagnosis of malaria. These studies have focused their attention to acquire images of smears of the specimen which can then be sent over a telecommunication channel for analysis.

Automated malaria parasite detection has been attempted by several groups using various approaches such as color histogram based thresholds, morphological operators, granulometry, shape and texture parameters, etc.

In ref. [6], the authors reported the study of analysis of blood cell images to detect abnormalities in which they binarized the original image using a fuzzy measure and labeled all the cells in the image. These labeled cells were then classified into red blood cells (RBC), white blood cells (WBCs) and platelets using a hierarchical neural network model with the help of their features, size and color.

In the work of Di Ruberto et al. [7, 8], digital images were acquired by scanning the color photographs of stained malarial blood from a microscope. They used morphological granulometry to obtain the size distribution of RBCs. They used HSV color space components for analysis. A simple thresholding of these components and their product generated a marker image for WBC and parasites. In this study, morphological approach was used to classify the parasites into immature trophozoites, mature trophozoites and gametocytes.

In the work of Ross et al. [9], digital images were acquired in JPEG format and were preprocessed. The complemented green component of the true color image was used in this study. The preprocessing and analysis steps followed the work of Di Ruberto et al. [7, 8] to some extent. This was followed by image segmentation, generation of features and classification. The preprocessing of the image included filtration by a median filter, followed by a morphological area closing filter. The overlapping RBCs were distinguished from the single ones with the help of granulometry. The overlapping RBCs were then separated using morphological operators used by Di Ruberto et al. [7, 8]. Geometric, texture features and color features were generated from the obtained RBCs. They used a two stage tree classifier with infection classified as positive or negative at the first node and the species assigned at the second node.

Sio et al. in [10] developed an image analysis based program for accurate determination of parasitemia called MalariaCount. The algorithm consists of four stages namely edge detection, edge linking, clump splitting and parasite detection. Adaptive histogram equalization was used for preprocessing.

Tek et al. in [11] reported a method to detect malaria parasites in which they separated stained and non-stained pixels using a Bayesian pixel classifier and then classified the stained pixels as parasites or non-parasites using a distance weighted K-nearest neighbor classifier. Four features namely color histogram, Hu moments, shape measurements and color autocorrelogram were selected for the second step of classification.

Diaz et al. [12] reported a color segmentation technique for separation of pixels into parasite, red blood cell and background, based on standard supervised classification algorithms. Supervised classification techniques namely KNN, Naive Bayes, SVM and Neural network were evaluated on different color spaces namely RGB, normalized RGB, HSV and YCbCr. For the classification of pixels into one of the three classes, training sample sets were manually extracted by an expert, and each of the pixels of the training sample was labeled accordingly. A classification model was trained for each of the color spaces. The classified color space was then used as a look-up table.

In the work of Le et al. [13], the authors report a semi-automatic image processing approach to detect malaria parasites. Here, the comparison based approach differentiated solid components in the smears by exploiting the inter-relations between different observations and radiometric representation. The algorithm is comprised of six steps namely nucleated components detection, image decomposition, erythrocyte size estimation, leukocytes and malarial gametocytes identification, erythrocyte segmentation and finally parasitemia estimation.

In Ref.[14], the authors provide an overview of computer vision studies of malaria diagnosis and provide critique of these works. They also provide a perspective of the future work for realization of automated microscopy diagnosis of malaria.

There is a need for a decision support system which aids a hematologist in malaria parasite detection in thin blood smear images. The management of review of all microscopic views of the sample is quite a laborious task for the hematologist. A decision support system is highly desirable in a laboratory setup of a hospital for regular use and also for screening programs where a large number of cases need to be managed in a short period of time. In this paper, we attempt to develop an algorithm to detect malaria parasites which can be used to develop a decision support system.

Methods

Blood specimen collected from patients is spread as thin blood smear, stained with Giemsa stain and examined with a 100× oil immersion objective. This study uses a total of 200 images out of which 100 images represented P. vivax infection, another 50 images represented P. falciparum infection and the other 50 images represented joint manifestation of P. vivax and P. falciparum infection. Out of the 200 images, 20 images were used to manually study the typical sizes of RBC, WBC, schizonts and the threshold values suitable for the segmentation. These images were carefully acquired to obtain uniform staining and illumination. All the procedures followed were in accordance with the ethical standards of the responsible committee on human experimentation of the university.

Results

The outputs of different steps of the algorithm and the classification parameters used are shown in Figs. 2 and 3, respectively.

This automated analysis was correlated with the manual evaluations, and the correlation was found to be highly significant. The statistical analysis was done with the help of freely available software named GraphPad. The Spearman correlation coefficient r was 0.888 with a confidence interval of 0.838 to 0.923, p < 0.0001.

The classification of regions with the use of two percentages of pixels representing the chromatin and ring as parameters is shown in Fig. 3. Blue circle represents normal RBC, and red star represents malaria parasite.

The details of parasites actually present, the number of parasites detected by the algorithm and the number of parasites which were missed by the algorithm are tabulated in Table 1.

The screenshots of the mobile application are shown in Figs. 4, 5 and 6. The composite image is displayed as an array of subimages, each subimage representing the affected RBC as detected by the image analysis program, shown in Fig. 4.

The expert can click on the subimage to view it individually, and in case he finds it as a false positive case, he can indicate so by selecting the radiobutton ‘No’ at the bottom of the screen, as shown in Fig. 5. This can be repeated for all suspicious subimages, as required by the expert. Once he is done with this step, he clicks on the ‘Submit confirmation’ button which takes him to a screen with the prescription option, shown in Fig. 6.

Also, the species of the parasite can be selected by the expert. The ‘send’ button generates a text message which indicates the false positive cases, species of the parasite, suggested drug and dosage and any other comments, and sends it to the laboratory. Another application on the receiver’s side can read the SMS from the expert, parse it and generate a prescription for the patient.

Discussion

The previous studies on blood smear analysis used traditional image processing techniques like thresholding [9], edge detection [10, 16], region growing [17], morphological operations and granulometry [7–9, 18] for the study of parasitemia. Some studies have focused on estimation of RBCs in case of overlapping cells [10, 16, 19]. Authors in [13] have developed a comparison based approach to handle the variations that may occur during sample preparation and image acquisition phase. The practicality, robustness, accuracy of these solutions and their applicability for diagnosis are questioned in [14]. Many of these efforts to automate malaria parasite detection fail to efficiently differentiate between parasites, artifacts and healthy blood components [8, 16, 20]. There are a few studies which differentiate between an infected RBC and a healthy RBC [9, 11]. Those few studies which address this problem [9, 11] fail to provide results comparable to manual analysis.

Malaria being a contagious disease, the primary goal of automation of the diagnosis is to provide quick and efficient treatment which can control the spread of malaria. Decision on treatment depends mainly on the correct identification of the species of the parasite, parasitemia and the dominant life stage of parasites in the specimen. Identification of the species of the parasite requires a lot of heuristic judgment, and any mistake in this step can have a very serious impact. A technique which attempts automation of this step has to be extremely robust and highly accurate.

Our approach attempts to address this issue. Malaria is prevalent in developing countries, and the microscopists are either overworked or not adequately trained. Mass screening requires reliable and fast technique to diagnose malaria. Moreover, facilities for diagnosis and treatment of malaria are not available in all rural or primary health centers. Studies on mobile phone based microscopy have been reported in [4, 5] to diagnose malaria. But it is not very practical to send images of all the views of the specimen to a remote expert over a telecommunication channel for diagnosis. An image analysis algorithm which can detect parasites with very high sensitivity and a facility for teleconsultation with an expert for the final decision on the treatment can handle malaria efficiently. Hence, we developed an algorithm which picks up all abnormal regions from all the views of the specimen and puts all of them into a composite image which can be sent over a telecommunication channel for an expert opinion. The identification of species and decision on treatment can be obtained by an expert. This facilitates accurate diagnosis and treatment.

The algorithm evaluates the proportion of colors present in a trophozoite, schizont and gametocyte and only if the region of interest satisfies the color contents specific to these, it is identified as a malaria parasite. Depending on how far the color proportions match, a color score is given to each of the abnormal regions selected. The score value for each region classified as infected indicates the confidence factor for the decision of classification. This technique reduces possibility of false positive detection and thus is more robust in handling WBCs, platelets and other staining artifacts.

We used MATLAB on a Pentium machine with 1 GB RAM. The computation of each image took around 2 min. In our approach, we used a gray-level threshold technique for background elimination. Alternately, the green channel could also be used to threshold and obtain the background as reported in [14]. We used morphological operations for RBC detection step. This is a time consuming task. This could also be achieved by using the initial selection of a typical RBC by the hematologist, which could provide the information of the approximate size of the RBC. This information could be used to set a threshold for the size of connected component. These steps can be adopted for achieving better computational efficiency.

The proposed method is designed to keep the false negative rate at a minimum. The false negative rate of this algorithm is about 7% and is tolerable. The false positive rate is about 20%. However, some refinements could improve the results of the algorithm, such as an improvement of the RBC detection sub-algorithm or a further observation for better threshold values for color segmentation and classification. For this work, we have used a carefully acquired set of images from a single laboratory. To make it more robust to color tone variations, a color normalization technique could also be included. As suggested in [14], a count of WBC could also be maintained to obtain an accurate value for parasitemia.

Conclusion

In this paper, we demonstrate the use of color image analysis to develop a decision support system for malaria parasite detection. A prototype of a mobile application has been developed to facilitate communication between a rural primary health centre which may not even have internet connectivity and a remote expert. This enables quick, efficient and easy way to manage malaria by facilitating accurate diagnosis and hence the correct treatment at the right time.