Application of principal component analysis to multispectral-multimodal optical image analysis for malaria diagnostics
© Omucheni et al.; licensee BioMed Central. 2014
Received: 19 September 2014
Accepted: 1 December 2014
Published: 11 December 2014
Multispectral imaging microscopy is a novel microscopic technique that integrates spectroscopy with optical imaging to record both spectral and spatial information of a specimen. This enables acquisition of a large and more informative dataset than is achievable in conventional optical microscopy. However, such data are characterized by high signal correlation and are difficult to interpret using univariate data analysis techniques.
In this work, the development and application of a novel method which uses principal component analysis (PCA) in the processing of spectral images obtained from a simple multispectral-multimodal imaging microscope to detect Plasmodium parasites in unstained thin blood smear for malaria diagnostics is reported. The optical microscope used in this work has been modified by replacing the broadband light source (tungsten halogen lamp) with a set of light emitting diodes (LEDs) emitting thirteen different wavelengths of monochromatic light in the UV–vis-NIR range. The LEDs are activated sequentially to illuminate same spot of the unstained thin blood smears on glass slides, and grey level images are recorded at each wavelength. PCA was used to perform data dimensionality reduction and to enhance score images for visualization as well as for feature extraction through clusters in score space.
Using this approach, haemozoin was uniquely distinguished from haemoglobin in unstained thin blood smears on glass slides and the 590–700 spectral range identified as an important band for optical imaging of haemozoin as a biomarker for malaria diagnosis.
This work is of great significance in reducing the time spent on staining malaria specimens and thus drastically reducing diagnosis time duration. The approach has the potential of replacing a trained human eye with a trained computerized vision system for malaria parasite blood screening.
Multispectral imaging is a novel imaging technology that integrates spatial and spectral information of an object under observation. Its use has been reported in food quality analysis [1–4], medical applications [5, 6] and remote sensing , among other applications. A multispectral image consists of a set of congruent grey-level images, each acquired at different wavelengths of light. The values of the grey levels (the recorded intensity value in each pixel) represent a measured intensity of the radiation used to capture the image. The number of grey level (intensity) images generally varies from four to a few tens depending on the spectral resolution required or equipment limitations. A typical multispectral image may consist of 512×512 pixels in the spatial domain, with each pixel (actually a voxel) consisting of a measurement spectrum for each of the wavelengths at which grey level imaging was done. As such, the amount of data obtained from a multispectral image is very large and may consist of correlated spectra. Multivariate chemometric tools such as principal component analysis (PCA) are handy in analysing such data because they search for latent structure in the raw data that best describe the measured variables to give desired information.
A chemometric-based method towards rapid malaria diagnosis in unstained thin blood smear was sought for development. Malaria is a serious, infectious disease caused by a parasitic protozoon known as Plasmodium. It is transmitted from one human being to another via saliva of a feeding female Anopheles mosquito. According to World Health Organization global estimates of malaria cases and deaths, an approximated 0.627 million of the 207 million people infected by the disease in the year 2012 succumbed to this menace, 90% being from Africa . Four species of Plasmodium cause malaria in humans: Plasmodium ovale, Plasmodium vivax, Plasmodium malariae, and Plasmodium falciparum. The latter is responsible for the highest mortality.
Examination of stained blood film under a light microscope for presence of Plasmodium is currently the gold standard for laboratory confirmation of malaria. In this technique, peripheral blood from a finger (or the heel in young infants) is smeared on a glass slide, stained and fixed (for thin blood smear) to highlight the malaria parasites. Giemsa staining is the most commonly used method . The staining process can take up to 30 minutes for thin film preparation while examination of 100 fields (a field is area that is visible in a microscope) in a thick blood smear (the film preparation method that is a little faster) has been found to miss infection by up to 20% . Moreover, this method is tedious as the microscopist or the examining technician has to examine as many red blood cells as possible on a glass slide with his/her own eye. A correct diagnosis is done only after attentive and careful observation of at least 100 microscopic fields has been examined in thick blood smear, and a number of morphological characteristics have been drawn to identify the species in thin blood smear. The time spent in the staining process and manual examination of many fields is the cause of delays in releasing laboratory results in hospitals. In addition, the results are subjective and may not be reproducible depending on the expertise of the examining personnel and the time spent on examination of the specimen. These limitations experienced in conventional light microscopy have stimulated research in development of alternative diagnostic methods that could be rapid, sensitive and affordable.
Most current malaria diagnostic techniques are based on any of the following five approaches: immunochromatographic tests for malaria antigen in blood [10, 11], polymerase chain reaction (PCR) [12–14], a variety of spectroscopic means of detection of haemozoin [15–17], and automated imaging microscopy of malaria parasites’ cell morphology and colour [18–20]. These methods are either accurate but very expensive, as in the case of PCR, or cheap but inaccurate due to limited information obtained from the sample, as in the case of colour imaging microscopy. In this work, PCA was combined with multispectral and multimodal imaging to detect malaria parasites in unstained human blood media for malaria diagnostics.
The multispectral-multimodal imaging microscope
The imaging system used in this study was a multimodal, multispectral, light emitting diodes (LED)-illuminated imaging microscope developed by Mikkel et al. and advanced by Aboma . The equipment consists of a commercial optical microscope (Brunel compound microscope) modified by replacing the conventional illumination light (white light) with monochromatic LED lighting system. Thirteen LED illuminations centred at 375 nm, 400 nm, 435 nm, 470 nm, 525 nm, 590 nm, 625 nm, 660 nm, 700 nm, 750 nm, 810 nm, 850 nm, and 940 nm are employed. This range covers ultra-violet (UV), visible and near infra-red (NIR) regions of the electromagnetic spectrum. The LEDs are mounted in a quasi-hemispherical container made of Teflon and illuminate same spot on a 5-mm Opal light diffuser (Edmund Optics, UK) to provide uniform illumination to the sample. To achieve multispectral imaging, the LEDs are activated sequentially through a computer controlled DAQ (National Instruments Inc, UK). For each LED illumination, a grey-scale image is recorded, thus resulting in 13 grey-scale images of same spot on the specimen.
The sample is illuminated in three different modes: bright field (transmission), dark field (scattering) and reflection. For transmission measurement, the sample is illuminated by a set of LEDs located directly below the sample. In reflection mode, light reaches the sample from above via a system of optics, which includes a cassegrain objective. The cassegrain objective (×15 Reflx™®, 0.28 Numerical Aperture objective, Edmund Optics, UK) is a reflective objective that is essential for reduction of chromatic aberration in all the three modes of imaging. Dark field imaging is achieved through the use of a ring coupled on an optical fibre. The ring provides an oblique illumination that cannot be accepted by objective’s aperture, hence only light which is scattered by the specimen is accepted by the objective. An image formed in this way has bright objects superimposed upon a dark background.
In vitro unstained thin blood smear malaria samples were obtained from Kenya Medical Research Institute (KEMRI). The samples had already been positively identified by an expert microscopist as infected and containing all the asexual stages of P. falciparum. For each sample, three measurements (transmission, reflectance and scattering) were made. For each measurement, the sample was focused such that the camera ‘sees’ the sample from all the three modes at same image plane. Since LEDs of different wavelengths have different illumination intensity and the camera quantum efficiency varies with wavelength, a protocol was created, at every measurement, in the software (developed in LabView-National Instruments Inc, UK) by adjusting the illumination intensity (current adjustment) of the LEDs, the exposure time and gain of the camera to obtain a good dynamic range of images without saturating the camera.
For each mode, three images of each wavelength were taken using the same protocol. These were the sample image, I(λ)s, reference image, I(λ)r and dark image, I(λ)d. For transmission measurements, I(λ)r is an image of an empty glass slide. In reflection and scattering modes, I(λ)r is the image of reflecting and scattering sides respectively of an Opal diffuser (PO:006208–00 SM-Edmund Optics) known to have an average reflectance value of 50%. The dark current image I(λ)d in transmission mode is taken by using the same transmission protocol as that of the sample but with the LEDs switched off. In reflection, I(λ)d is obtained using the same protocol as that of the reflection sample image with nothing on the light path (sample and slide removed) while in dark field mode I(λ)d is obtained with an empty glass slide using scattering sample image protocol.
where I(λ)s is sample image, I(λ)r reference image, and I(λ)d is dark image. The subtraction is done on a pixel-by-pixel basis for each wavelength (λ). Image correction and the subsequent processing and analysis were performed using MATLAB® (MathWorks Inc, USA) software. A code was written to import the raw images stored in unsigned 16-bit representation, convert them to double class representation and apply the correction using Equation 1 in readiness for further analysis.
Where I(λ)d is the dark image normalized to unity.
Results and discussion
Identification of infected red blood cells in dark field
Since uninfected red blood cells are homogenous only their edges are seen to have scattered light. The whole scenario is observed against a dark background representing regions of little or no scattering.
However, identifying a red blood cell based on dark-field microscopic examination was not conclusive owing to the fact that morphological features well known to microscopists such as rings could not be resolved in the image. As such, it was difficult to separate a true positive from a false positive (such as an artefact, e.g., a dust particle) in the thin blood smear. Thus, a red blood cell identified as a positive in dark field had to be subjected to further scrutiny in order to arrive at a conclusion that it is a true positive. It was necessary to study the spectral dimension of images of both infected and non-infected red blood cells as identified in dark field.
Spectral response of Plasmodium in blood
where Io is the incident intensity, It the transmitted intensity and A the absorbance.
It is important to note that these spectra are not absolute but only illustrate the main spectral features. The absorption spectra of infected and non-infected red blood cells have a common intense absorption band centred near 400 nm, which is due to electronic transitions of the porphyrin ring found in haem. However, there are slight differences in the spectra: a healthy red blood cell has slightly higher absorbance in the 470–590 nm wavelength range, whereas an infected one has slightly higher absorbance in the 625–700 nm spectral range.
Plasmodium, the malaria parasite, is a complete unicellular organism with cell membrane, nucleus and myriad organelles. On the other hand, a mature red blood cell consists mainly of haemoglobin protein without nucleus. The invasion of the red blood cell by Plasmodium with its cellular complexity presents some spectral response that requires a biochemical understanding of the precise origin of the spectra. Malaria parasites have nucleic acid material, which is generally transparent in the visible region. In fact, nucleic acids are identified by their characteristic intense absorption at 260 nm , which is outside the range of the microscope used in this work. This therefore rules out any possibility of detecting the parasites’ DNA using the multispectral microscope in its current spectral range.
An attack by blood-feeding organisms, such as malaria parasites, is associated with production of a large amount of waste product known as haemozoin (also known as malaria pigment). Haemozoin is a haem crystal produced by Plasmodium in order to detoxify free haem released when haemoglobin is digested in the digestive vacuoles of the parasites. Free haem is a potent cytotoxic agent to the parasites because it promotes lysis of the cellular organisms and generates reactive oxygen species (ROS), which damage many biomolecules . Both haemoglobin and haemozoin contain tightly bound metal-ion-containing molecules known as prosthetic groups. The spectral signatures of prosthetic groups are known to be found in the visible region of the electromagnetic spectrum . Haemozoin is, therefore, of central importance in identifying infected red blood cells spectroscopically in unstained thin blood smear using the LED multispectral imaging microscope working in 375–940 nm range. The slightly higher absorbance of infected red blood cell in the spectral range 625–700 nm but lower absorbance in the 470–590 nm can therefore be attributed to reduction of haemoglobin but increase in haemozoin in an infected red blood cell in the trophic stage of the parasite.
Reflectance spectra for both parasite-infected and non-infected red blood cells were also generated from reflectance images. However the differences between infected and uninfected red blood cell spectra were not reproducible, probably due to misalignment of the instrument and hence are not reported here.
Image analysis based on multivariate chemometric techniques
The spectral data generated from the images were explored for any observable relational phenomena, such as clusters of pixels, gradients between the clusters and outliers. A good starting point was to use PCA, a technique that has been extensively employed successfully in multivariate image applications [26–31] to segment features based on their spectral characteristics.
Segmentation using principal component analysis (PCA)
From Figures 5, 6 and 7, it can be seen that the pixels from the parasite form a small cluster within the pixel cluster of the red blood cell. Overlapping of these clusters may have arisen from the fact that pixels in an infected red blood cell image contain mixed spectra of haemoglobin and haemozoin. In addition, the spectral features of haemozoin and haemoglobin are highly correlated. Both haemoglobin and haemozoin contain haem (a prosthetic group), which exists as a large heterocyclic organic ring called porphyrin with an iron ion at the centre. In haemozoin, this ring exists as ferriprotoporphyrin IX  whereas the non-protein part of haemoglobin exists as ferroprotoporphyrin IX. These two compounds exhibit π-π* electronic transitions of the porphyrin ring visualized as the strong absorption band (Soret band) centred near 400 nm as seen in Figure 3. Their differences, however, arise from different locations of their vibronic transitions which occur at 540 nm and 575 nm for oxy-haemoglobin and 650 nm for haemozoin.
Important spectral band selection
Image generation by selective illumination of red blood cells
In the image captured at 625 nm (Figure 10), the parasite region appears brighter due to its higher absorption in the image in comparison with other parts of red blood cells (see the grey-level scale). However, in the image captured at 400 nm (Figure 9) and that captured at 940 nm (Figure 11), there is little intensity difference between the parasites and the red blood cells. The reason for this is that at 400 nm and at 940 nm, the malaria pigment has almost identical absorption properties to those of haemoglobin. At 625 nm (and at the other three important spectral bands viz 590 nm, 660 nm and 700 nm) however, there are slight differences in absorption of haemoglobin and haemozoin, making the parasite more visible.
These results are in agreement the findings of Aboma Merdasa et al. . According these findings, the greatest difference between infected and uninfected red blood cells when analyzed using singular value decomposition and hierarchical clustering for all the three modes of imaging is found at around 630 nm.
The application of PCA to multispectral image analysis has been demonstrated. The spectral images were recorded using a simple LED-illuminated microscope equipped with a CMOS camera. Using PCA malaria parasites in infected red blood cells were successfully segmented and important spectral bands for optical imaging of malaria parasites in unstained thin blood smear identified. This is of great significance in reducing the time spent on staining malaria specimens and thus drastically reducing diagnosis time duration.
We would like to thank International Science Programme (ISP) for their financial support on this project and African Spectral Imaging Network (AFSIN) for providing technical assistance.
- Wold JP, Westad F, Heia K: Detection of parasites in cod fillets by using SIMCA classification in multispectral images in the visible and NIR region. Appl Spectrosc. 2001, 55: 1025-1034. 10.1366/0003702011952929.View ArticleGoogle Scholar
- Noordam JC, Broek WH, Buydens LMC: Detection and classification of latent defects and diseases on raw French fries with multispectral imaging. J Sci Food Agri. 2005, 85: 2249-2259. 10.1002/jsfa.2226.View ArticleGoogle Scholar
- Aleixos N, Blasco J, Navarron F, Molto E: Multispectral inspection of citrus in real-time using vision and digital signal processors. Comput Electron Agric. 2002, 33: 121-137. 10.1016/S0168-1699(02)00002-9.View ArticleGoogle Scholar
- Polder G, Heijden GW, Young IT: Spectral image analysis for measuring ripeness of tomatoes. Transact ASAE. 2002, 45: 1155-1161.View ArticleGoogle Scholar
- Zoueu TJ, Loum GL, Haba CT, Mikkel B, Menan H: Optical microscope based on multispectral imaging applied to Plasmodium diagnosis. J Appl Sci. 2008, 8: 2711-2717. 10.3923/jas.2008.2711.2717.View ArticleGoogle Scholar
- Ihor VB, Svitlana YB: Fast multi-spectral imaging technique for detection of circulating endothelial cells in human blood samples. J Biomed Opt. 2012, 17: 081404–1-Google Scholar
- Goetz AF, Vane G, Solomon JE, Rock BN: Imaging spectrometry for earth remote sensing. Science. 1985, 228: 1147-1153. 10.1126/science.228.4704.1147.View ArticlePubMedGoogle Scholar
- World Health Organization: World Malaria Report 2013. 2013, Geneva: World Health OrganizationGoogle Scholar
- Craig MH, Sharp BL: Comparative evaluation of four techniques for the diagnosis of Plasmodium falciparum infections. Trans R Soc Trop Med Hyg. 1997, 91: 279-282. 10.1016/S0035-9203(97)90074-2.View ArticlePubMedGoogle Scholar
- Garcia M, Kirimoama S, Marlborough D, Leafasia J, Rieckmann KH: Immunochromatographic test for malaria diagnosis (Letter). Lancet. 1996, 347: 1549-View ArticlePubMedGoogle Scholar
- Shiff CJ, Minjas JN, Premji Z: The Parasight®-F test: a simple rapid manual dipstick test to detect Plasmodium falciparum infection. Parasitol Today. 1994, 10: 494-495. 10.1016/0169-4758(94)90164-3.View ArticlePubMedGoogle Scholar
- Oliveira DA, Shi YP, Oloo AJ: Field evaluation of polymerase-chain reaction based nonisotopic liquid hybridization assay for malaria diagnosis. J Infect Dis. 1996, 173: 1284-1287. 10.1093/infdis/173.5.1284.View ArticlePubMedGoogle Scholar
- Snounou G, Viriyakosol S, Jarra W, Thaithong S, Brown KN: Identification of the four human malaria parasites in field samples by the polymerase-chain-reaction and detection of a high prevalence of mixed infections. Molec Biochem Parasitol. 1993, 58: 283-292. 10.1016/0166-6851(93)90050-8.View ArticleGoogle Scholar
- Tirasophon W, Rajkulchai P, Ponglikitngkol M, Wilairat P, Boonsaeng V, Panyim S: A highly sensitive, rapid and simple polymerase-chain-reaction method to detect human malaria (Plasmodium falciparum and Plasmodium vivax) in blood samples. Am J Trop Med Hyg. 1994, 51: 308-313.PubMedGoogle Scholar
- Belisle JM, Costantino S, Leimanis ML, Bellemare MJ, Bohle DS, Georges E, Wiseman PW: Sensitive detection of malaria infection by third harmonic generation imaging. Biophys J Biophys Lett. 2007, 94: L26-L28.View ArticleGoogle Scholar
- Ong CW, Shen ZX, Ang KH, Kara UAK, Tang SH: Raman microspectroscopy of normal erythrocytes and Plasmodium berghei-infected erythrocytes. Appl Spectrosc. 2002, 56: 1126-1131. 10.1366/000370202760295340.View ArticleGoogle Scholar
- Grant TW, Leann T, Samantha D, Don M, Bayden RW: Resonance Raman spectroscopy can detect structural changes in haemozoin (malaria pigment) following incubation with chloroquine in infected erythrocytes. FEBS Lett. 2008, 582: 1087-1092. 10.1016/j.febslet.2008.02.062.View ArticleGoogle Scholar
- Minh TL, Timo RB, Claudia K, Preiser PR: A novel semi-automatic image processing approach to determine Plasmodium falciparum parasitemia in Giemsa-stained thin blood smears. BMC Cell Biol. 2008, 28: 9-15.Google Scholar
- Ross NE, Pritchard CJ, Rubin MD, Adriano GD: Automated image processing method for the diagnosis and classification of malaria on thin blood smears. Med Biol Eng Compt. 2006, 44: 427-436. 10.1007/s11517-006-0044-2.View ArticleGoogle Scholar
- Gloria D, Fabio AG, Romero E: A semi-automatic method for quantification and classification of erythrocytes infected with malaria parasites in microscopic images. J Biomed Inform. 2009, 42: 296-307. 10.1016/j.jbi.2008.11.005.View ArticleGoogle Scholar
- Brydegaard M, Guan Z, Svanberg S: Broad-band multispectral microscope for imaging transmission spectroscopy employing an array of light-emitting diodes. Am J Phys. 2008, 77: 104-110.View ArticleGoogle Scholar
- Merdasa A: Multispectral Microscopy With Applications to Malaria Detection. Master’s Thesis. 2010, Division of Atomic physics, Lund UniversityGoogle Scholar
- Jamjoom GA: Dark-field microscopy for detection of malaria in unstained blood films. J Clin Microbiol. 1983, 17: 717-721.PubMed CentralPubMedGoogle Scholar
- Hammes G: Spectroscopy for the Biological Sciences. 2005, New Jersey: John Wiley & Sons LtdView ArticleGoogle Scholar
- Oliveira MF, Gandara P, Braga MS, Silva RJ, Mury BF, Marilvia D, Diego M, Marcos A, Pedro O: Heme crystallization in the midgut of triatomic insects. Comp Biochem Physiol C. 2006, 146: 168-174.Google Scholar
- Shah S, Thigpen J: Multispectral Microscopy for Cell Differentiation in Thyroid Cytology. Proceedings of IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems: 20–22 Aug. 2008. 2008, Seoul: IEEE, 267-271.Google Scholar
- Kvaal K, Wold JP: Mapping lipid oxidation in chicken meat by multispectral imaging of autofuorescence. Appl Spectrosc. 2000, 54: 900-909. 10.1366/0003702001950300.View ArticleGoogle Scholar
- Gowen AA, Donnell CP, Taghizadeh M, Cullen PJ, Frias JM, Downey G: Hyperspectral imaging combined with principal component analysis for bruise damage detection on white mushrooms (Agaricus bisporus). J Chemometr. 2007, 22: 259-267.View ArticleGoogle Scholar
- Ariana DP, Lu R, Guyer DE: Near-infrared hyperspectral reflectance imaging for detection of bruises on pickling cucumbers. Comput Electron Agric. 2006, 53: 60-70. 10.1016/j.compag.2006.04.001.View ArticleGoogle Scholar
- Costa JC, Alves MM, Ferreira EC: Principal component analysis and quantitative image analysis to predict effect of toxics in anaerobic granular sludge. Bioresour Technol. 2009, 100: 1180-1185. 10.1016/j.biortech.2008.09.018.View ArticlePubMedGoogle Scholar
- Baronti S, Casini A, Lotti F, Porcinai S: Principal component analysis of visible and near-infrared multispectral images of works of art. Chemometr Intell Lab Syst. 1997, 39: 103-114. 10.1016/S0169-7439(97)00047-6.View ArticleGoogle Scholar
- Ray DJ, Huete AR: Interpreting vegetation indices. Prev Vet Med. 1991, 11: 185-200. 10.1016/S0167-5877(05)80004-2.View ArticleGoogle Scholar
- Coy DF, Phitsamai K: The state of ferriprotoporphyrin IX in malaria pigment. Biol Chem. 1987, 262: 15552-15555.Google Scholar
- Merdasa A, Brydegaard M, Svanberg S, Zoueu JT: Stain-free malaria diagnostics by multispectral and multimodality light-emitting-diode microscopy. J Biomed Opt. 2013, 18: 036002-10.1117/1.JBO.18.3.036002.View 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/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.