Reliable enumeration of malaria parasites in thick blood films using digital image analysis
© Frean; licensee BioMed Central Ltd. 2009
Received: 22 July 2009
Accepted: 23 September 2009
Published: 23 September 2009
Quantitation of malaria parasite density is an important component of laboratory diagnosis of malaria. Microscopy of Giemsa-stained thick blood films is the conventional method for parasite enumeration. Accurate and reproducible parasite counts are difficult to achieve, because of inherent technical limitations and human inconsistency. Inaccurate parasite density estimation may have adverse clinical and therapeutic implications for patients, and for endpoints of clinical trials of anti-malarial vaccines or drugs. Digital image analysis provides an opportunity to improve performance of parasite density quantitation.
Accurate manual parasite counts were done on 497 images of a range of thick blood films with varying densities of malaria parasites, to establish a uniformly reliable standard against which to assess the digital technique. By utilizing descriptive statistical parameters of parasite size frequency distributions, particle counting algorithms of the digital image analysis programme were semi-automatically adapted to variations in parasite size, shape and staining characteristics, to produce optimum signal/noise ratios.
A reliable counting process was developed that requires no operator decisions that might bias the outcome. Digital counts were highly correlated with manual counts for medium to high parasite densities, and slightly less well correlated with conventional counts. At low densities (fewer than 6 parasites per analysed image) signal/noise ratios were compromised and correlation between digital and manual counts was poor. Conventional counts were consistently lower than both digital and manual counts.
Using open-access software and avoiding custom programming or any special operator intervention, accurate digital counts were obtained, particularly at high parasite densities that are difficult to count conventionally. The technique is potentially useful for laboratories that routinely perform malaria parasite enumeration. The requirements of a digital microscope camera, personal computer and good quality staining of slides are potentially reasonably easy to meet.
Microscopic examination of Giemsa-stained blood films is widely relied upon for routine malaria diagnosis and particularly, for parasite density quantitation. Estimation of parasite burden in falciparum malaria is done for several reasons: as an indicator of risk of severe and complicated disease, especially in non-immune patients ; as a measure of response to treatment; as an aid to clinical decision-making about the likely cause of febrile illness in highly endemic areas ; and as an end-point in clinical trials of anti-malarial drugs or vaccines, at a pre-determined parasite density threshold .
It has been pointed out that there is a striking lack of evidence to support widely-held assumptions about the accuracy and consistency of malaria microscopy [2, 4]. Studies have shown substantial intra- and inter-observer inconsistencies in density quantitation [2, 5], and experience in proficiency testing of malaria microscopy in Africa supports this [6, 7]. As counting parasites is tedious and tiring for microscopists, automation in the form of digital image analysis is an obvious potential solution . Recent reports have described progress in applying this technology to thin film parasitaemia estimation [9–11]. However, despite some inherent limitations , thick blood films are more generally used for determining parasite density .
This contribution describes proof-of-principle of a simple, low-cost image analysis technique that is highly effective in enumerating moderate to high malaria parasite densities in thick blood films. Specific aims of this project were to develop a process to minimize user intervention, to avoid custom-written software or specialized hardware, and to thereby make the technique readily accessible to suitably-resourced laboratories.
Thick blood films, Giemsa-stained to uniform standards  were obtained from our slide bank of malaria proficiency-testing specimens, or loaned from a similar collection (K. Lilley, Army Malaria Institute, Brisbane). Blood samples for malaria microscopy proficiency testing were collected and used with ethical approval of the Human Research Ethics Committee (Medical), University of the Witwatersrand, Johannesburg (protocol number M051126). Twenty films containing Plasmodium falciparum were selected to provide parasite densities ranging from 5,000 to 500,000 parasites/μl. These included eight re-sampled or duplicate slides to test reproducibility of methods both within and between films prepared from the same blood specimens. Parasite densities had been previously established by experienced microscopists using conventional counting methods [7, 13]; namely, by counting parasites on thick films per 200 (or, in the case of very low densities, 500) leukocytes, multiplied by the patient's own leukocyte count, or if this was not available, a standard count of 8,000 leukocytes/μl. Specimens with very high counts (> 100 parasites per 100× objective field) were also counted on thin blood films as the proportion of infected erythrocytes multiplied by either the patient's red cell count, or, if this was not known, a standard red cell count (5 × 106 cells/μl).
Using a 50× objective in a conventional laboratory microscope (Olympus BX 41, Olympus Australia, Oakleigh, Victoria), sequential blood film images were captured by means of a Nikon DXM1200 digital camera  and Nikon ACT-1 software (Nikon Corporation, Tokyo, Japan) as uncompressed tagged image file format (TIFF) files at a resolution of 1,280 × 1,024 pixels. Apart from avoiding the irregular edges of the thick film and ensuring no overlapping of images, no special selection of captured fields was done. The number of leukocytes per image was recorded manually at the time of capture.
ImageJ (version 1.41), an open-access Java-based image-processing programme, was used for image analysis. In essence, the programme segments or classifies particles to be counted on the basis of their relative density (darkness) compared with the background, via a thresholding process. Particle size (area) and degree of roundness are other classification variables. Fine morphological and differential staining characteristics of parasites are ignored. Therefore, non-parasite particles, that is, artifacts of various types, may also be segmented and are collectively termed noise (N). The target particles (malaria parasites) are the signal (S).
Precise enumeration of parasites per image (the 'gold standard' for this study), was done by manually counting parasites on the captured images (in total, about 98,000 parasites in 497 images from 20 specimens were counted). A 'Point Picker' plugin  that digitally tags each counted parasite and records its coordinates for future reference, was used to facilitate manual counting. Using the particle analysis commands of ImageJ, parasites were then counted digitally on the same images. Between 20 and 30 (mean, 25) images per slide were analysed simultaneously in a stack; the amount of virtual memory available to the image analysis software constrains stack size. Three hundred images, representing 12 different slides (calibrators), were used in the calibration experiments described below; 197 new images from eight re-sampled or duplicate slides were used to validate the findings and assess reproducibility.
Statistical evaluations were done using Statistica 8.0 (StatSoft, Tulsa, OK). Because of non-normal distributions of data sets and small sample sizes (n < 30), non-parametric tests were used. Statistical evaluation at individual slide level was by signed rank tests that compared numerical results of manual and digital counting methods, together with the rank order correlation coefficient (R) as a measure of reliability of digital counts. Non-parametric ANOVA was used to compare collective counts by the three methods (conventional, manual, and digital). World Health Organization (WHO) criteria for evaluating accuracy of counts done by working microscopists, expressed as the percentage absolute discrepancy between experimental and reference counts, were used in a less stringent but practical comparison system .
During initial experiments to obtain a standard particle analysis algorithm applicable to thick film images, it was apparent that more user intervention, in terms of adjusting various algorithm variables, was required for films with lower parasite densities and/or later-stage trophozoites. Ultimately, a single adjustment factor was identified that would accommodate nearly all densities and sizes of parasites. This factor, designated RN, is the radius of the area used in the 'Remove outliers' command, 'Noise' submenu, 'Process' menu, of ImageJ. The command replaces a pixel by the median of the pixels in the surrounding area if it deviates from the median by more than a certain threshold value; the effect is to smooth irregular shapes and reduce non-specific noise, with increasing rigor as RN increases. Above a certain value of RN, signal is also removed. The effect of this on counts is variable and depends on the relative proportions of signal and noise.
Comparison of manual, digital, and conventional parasite density estimations
Number of parasites counted per slide
Parasite density, as parasites/μl
Comparison of manual and digital counts
Number of parasites counted per slide
Spearman correlation R; p
Wilcoxon test p
0.99; < 0.05
0.98; < 0.05
0.96; < 0.05
0.94; < 0.05
0.56; < 0.05
0.76; < 0.05
0.62; < 0.05
0.91; < 0.05
0.98; < 0.05
0.61; < 0.05
0.91; < 0.05
0.99; < 0.05
0.69; < 0.05
0.85; < 0.05
0.88; < 0.05
0.99; < 0.05
0.94; < 0.05
0.86; < 0.05
0.97; < 0.05
Comparison of digital parasite density estimations on pairs of duplicate or re-sampled* slides
Parasite density, as parasites/μl
Errors in parasite density estimation by conventional microscopy are common, and apart from possibly deleteriously influencing the management of individual patients, have the potential to produce major consequences for clinical efficacy trials of malaria vaccines or prophylactic drugs [6, 17]. Utility of automated digital particle analysis for enumerating parasites has previously been limited by the variability of size, shape, and staining characteristics of asexual malaria parasites on conventional stained thick blood films. In this study, highly accurate manual counts of a range of parasite densities made it possible to experiment extensively with digital counting methods, and to critically evaluate particle analysis algorithms. In addition, the accuracy of conventional counting methods applied to the same slides was assessed.
With ordinary image analysis particle counting, it is straightforward to adjust algorithm variables to achieve highly accurate counts on individual slides when the target (manual) count is known. However, this is clearly not relevant to routine parasite density determination where a set target is absent, and adjusting algorithm variables subjectively introduces unknown biases into counts. Variation in parasite size and shape preclude use of a universal algorithm, with or without additional manual adjustment. The method described here circumvents this problem by providing semi-automatic adjustment of the most important variable controlling the S/N ratio, based on certain frequency distribution parameters (kurtosis and skewness) of the particles being analysed. It can be seen that there are no decisions required from the user that might introduce subjectivity or bias.
Generally, digital counts correlated well with both manual and conventional counts (Figures 6 and 7), but conventional counts were significantly lower than digital and manual counts. In conventional counting of relative numbers of parasites and leukocytes, human operator biases, which are absent in the digital and manual counts described here, presumably account for this tendency to underestimate parasite densities. The method of digital counting described here essentially solves the problem of counting parasites at medium to high densities, but in images with low absolute numbers of parasites (< 140 parasites per 25 images), there is evidence that lower S/N ratios, because of misclassification, constrain the accuracy of counts. It may be possible to predict when digital counts are likely to be unreliable; in this data set it appears that when the quotient of the digital parasite count and RN is less than 20, a suboptimal S/N ratio is likely, despite the fact that the total count may be within acceptable limits. Further experiments will be required to verify this finding. Generally, lower densities are technically easier to count conventionally than high ones and, therefore, this restriction is not a practical problem (notwithstanding error due to sample distribution effects at low densities).
With regard to reproducibility of digital counts within and between films made from the same blood specimens, the data set is small, but simultaneous comparison with manual counts suggests that differences in digital counts in this subset of slides were mainly due to real inter-sample variation, with only a small contribution from counting errors.
Cost and availability of equipment may constrain application of image analysis; a digital camera-equipped microscope and a computer are required. Another limitation of the technique is that stain and other artifacts may contribute to unwanted noise. It follows that good-quality staining with minimal residual stain precipitate is necessary for optimum results. Visual inspection of smears during image capture should alert the microscopist to significant numbers of Howell-Jolly bodies, yeasts, or other particles that may also occasionally contribute to noise. Dim images tend to be noisy; it is important to optimize the brightness and contrast of captured images for automatic thresholding by the image analysis software. Another constraint is the need to capture substantial numbers of digital images of the specimens, which is not difficult but is time-consuming, taking up to an hour to carefully capture 25 images. Computer-controlled motorized microscope stages and automatic focusing are solutions to this problem that are already available from some manufacturers, but would add to costs. In contrast, the subsequent digital counting process is fast, requiring about 5 to 10 minutes per specimen to complete.
This proof-of-principle study has shown that it is possible to achieve high standards of accuracy and reproducibility of thick film malaria parasite counts by digital image analysis. Digital counts were generally well within the acceptable limits of accuracy recommended by the WHO . High quality free software and semi-automation of the counting process make this technique potentially widely accessible to many diagnostic laboratories. Further evaluation of the method using a larger number of malaria blood films and different digital camera systems, is the next objective. The principle may suggest a general solution for automated biological particle counting, with minimal operator input required, when some variation in size and shape of the target is present.
Grateful thanks go to Rita van Deventer, Leigh Dini, and Bhavani Poonsamy for the blood films and conventional counts; to Dennis Shanks, Ken Lilley, Bob Cooper and staff of AMI for use of facilities and equipment; and to the National Institute for Communicable Diseases and the School of Pathology of the University of the Witwatersrand and National Health Laboratory Service, for providing opportunity and financial support.
- World Health Organization: Guidelines for the treatment of malaria. 2006, Geneva: WHOGoogle Scholar
- O'Meara WP, McKenzie FE, Magill AJ, Forney JR, Permpanich B, Lucas C, Gasser RA, Wongsrichanalai C: Sources of variability in determining malaria parasite density by microscopy. Am J Trop Med Hyg. 2005, 73: 593-598.PubMed CentralPubMedGoogle Scholar
- Bejon P, Lusingu J, Olotu A, Leach A, Lievens M, Vekemans J, Mshamu S, Lang T, Gould J, Dubois MC, Demoitié MA, Stallaert JF, Vansadia P, Carter T, Njuguna P, Awuondo KO, Malabeja A, Abdul O, Gesase S, Mturi N, Drakeley CJ, Savarese B, Villafana T, Ballou WR, Cohen J, Riley EM, Lemnge MM, Marsh K, von Seidlein L: Efficacy of RTS, S/AS01E vaccine against malaria in children 5 to 17 months of age. NEJM. 2008, 359: 2521-2532. 10.1056/NEJMoa0807381.PubMed CentralView ArticlePubMedGoogle Scholar
- Bejon P, Andrews A, Hunt-Cooke A, Sanderson F, Gilbert SC, Hill AVS: Thick blood film examination for Plasmodium falciparum malaria has reduced sensitivity and underestimates parasite density. Malar J. 2006, 5: 104-10.1186/1475-2875-5-104.PubMed CentralView ArticlePubMedGoogle Scholar
- O'Meara WP, Barcus M, Wongsrichanalai C, Muth S, Maguire JD, Jordan RG, Prescott WR, McKenzie FE: Reader technique as a source of variability in determining malaria parasite density by microscopy. Malar J. 2006, 5: 118-10.1186/1475-2875-5-118.PubMed CentralView ArticlePubMedGoogle Scholar
- Dini L, Frean J: Quality assessment of malaria laboratory diagnosis in South Africa. Trans R Soc Trop Med Hyg. 2003, 97: 675-677. 10.1016/S0035-9203(03)80101-3.View ArticlePubMedGoogle Scholar
- World Health Organization: Policy and Procedures of the WHO/NICD Microbiology External Quality Assessment Programme in Africa. 2007, Geneva: WHOGoogle Scholar
- Frean J: Improving quantitation of malaria burden with digital image analysis. Trans R Soc Trop Med Hyg. 2008, 102: 1062-1063. 10.1016/j.trstmh.2008.04.017.View ArticlePubMedGoogle Scholar
- Proudfoot O, Drew N, Scholzen A, Xiang S, Plebanski M: Investigation of a novel approach to scoring Giemsa-stained malaria-infected thin blood films. Malar J. 2008, 7: 62-PubMed CentralView ArticlePubMedGoogle Scholar
- Le MT, Bretschneider TR, Kuss C, Preiser PR: A novel semi-automatic image processing approach to determine Plasmodium falciparum parasitaemia in Giemsa-stained thin blood smears. BMC Cell Biol. 2008, 9: 15-10.1186/1471-2121-9-15.PubMed CentralView ArticlePubMedGoogle Scholar
- Sio SW, Sun W, Kumar S, Bin WZ, Tan SS, Ong SH, Kikuchi H, Oshima Y, Tan KSW: MalariaCount: an image analysis-based program for the accurate determination of parasitemia. J Microbiol Methods. 2007, 68: 11-18. 10.1016/j.mimet.2006.05.017.View ArticlePubMedGoogle Scholar
- World Health Organization: Informal consultation on quality control of malaria microscopy. 2006, Geneva: WHOGoogle Scholar
- World Health Organization: Basic malaria microscopy. 2001, Geneva: WHOGoogle Scholar
- Nikon DXM1200 camera technical details. [http://www.microscopedealer.com/products/imagearchiving/nikon_dxm1200.php]
- ImageJ. [http://rsb.info.nih.gov/ij/index.html]
- Point Picker plugin, ImageJ. [http://bigwww.epfl.ch/thevenaz/pointpicker/]
- O'Meara WP, Hall BF, McKenzie FE: Malaria vaccine efficacy: the difficulty of detecting and diagnosing malaria. Malar J. 2007, 6: 36-10.1186/1475-2875-6-36.PubMed CentralView ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.