Towards a needle-free diagnosis of malaria: in vivo identification and classification of red and white blood cells containing haemozoin

Smears were prepared from blood infected with two different species of Plasmodium, the rodent-infecting species P. yoelii and the human-infecting species P. falciparum. Plasmodium yoelii preferentially infects reticulocytes, similar to Plasmodium vivax, the predominant human-infecting species causing recurring malaria [15]. Blood from P. yoelii-infected mice was collected by cardiac puncture into heparinized tubes and used to prepare thin film smears. Unstained and unfixed thin film blood smears infected with P. falciparum at 0.5% parasitaemia were acquired from McGill University J.D. MacLean Centre for Tropical Diseases. Negative control smears were prepared from blood acquired from healthy normal volunteers. All smears were fixed in methanol and allowed to dry. A drop of liquid mounting medium containing DAPI fluorescent nuclear stain (Invitrogen P-36931) was applied to each smear prior to sealing with a coverslip. DAPI has been previously demonstrated as an effective contrast agent for parasite nuclei within infected red blood cells and for white blood cell differentiation [16]. Additionally, the DAPI absorbance peak, centered at 358 nm, does not interfere with Hz absorbance measurements.

To simulate WBC phagocytosis of Hz, isolated Hz was added to whole blood from a healthy volunteer using methods adapted from Frita et al. [17]. Hz was isolated from P. falciparum culture (Strain 3D7, acquired from PATH, Seattle, Washington). The cultured sample was washed with 1× phosphate buffered saline to remove glycerin and resuspended at 5% haematocrit, 5% parasitaemia. To induce RBC lysis, the sample was centrifuged for 10 min at 1000 g, resuspended in 1 ml of 0.15% saponin (Sigma), and incubated for 10 min on ice with intermittent vortexing. Following incubation, the sample was washed twice with phosphate buffered saline at 10,000 g for 15 min. The final isolated Hz pellet was resuspended in 1 ml of phosphate buffered saline.

Whole blood from a healthy volunteer was collected into heparinized tubes and diluted 1:1 in cell culture media (Sigma, RPMI 1640) at 10% FBS. Isolated Hz was added to the diluted blood 1:1 and the mixture was plated into a 48 well plate and incubated at 37 °C at 5% CO₂. After 4 h of incubation, samples were used to prepare thin film blood smears for analysis. A DAPI-containing mounting medium was applied to the smears before sealing with a coverslip.

Images of blood smears were acquired with an upright conventional microscope (Zeiss, Z1) using a 63× oil immersion objective. Two images were collected from each field of view, the first under brightfield transmission illumination and the second using a DAPI fluorescence filter set (Zeiss, Filter Set 49). For brightfield imaging, light from a broadband halogen lamp was passed through a narrow bandpass filter (Chroma, HQ 655/40) to isolate Hz absorbance and minimize haemoglobin Q-band absorbance.

Images were analysed using the open source platform ImageJ (National Institutes of Health, v1.47t), following the process illustrated in Fig. 1. Individual red and white blood cells were manually segmented using the brightfield images. A background intensity threshold (I_T), set as the average minimum brightfield pixel intensity of uninfected red blood cells, was applied to brightfield images to isolate Hz signal. Cell type was determined by visual examination of morphology in the corresponding DAPI image. WBCs were identified on the basis of their large nucleus observed under DAPI staining. RBCs have no DAPI signal, whereas parasite nuclei can be clearly distinguished in iRBCs. Parasite maturity was also determined using the DAPI signal, where iRBCs containing only a single parasite nucleus were identified as young trophozoites. Multinucleated iRBCs were labelled as late trophozoites. As a negative control, the same process was repeated for RBCs and WBCs in blood smears from non-infected mice and uninfected healthy volunteer human subjects.

Hz area and mean pixel intensity were recorded for each segmented cell. These data were used to calculate the Hz effective diameter (d_eff) and relative intensity difference (RID), two parameters assessed for their ability to discriminate between iRBCs and pWBCs. The Hz d_eff was estimated from the measured Hz area (A), where d_eff = 2 (A/π)^1/2. RID was measured as the percent difference between the background threshold intensity and the mean Hz pixel intensity, expressed as RID = (I_T − <I_HZ>)/I_T. Hz d_eff and RID were plotted for all cells versus cell type using the DAPI-determined cell type. Box plots characterizing the first quartile, median, and third quartile were generated for each cell type. Whiskers of these plots extend to ± 1.58 (IQR/(n^1/2)), where IQR is the interquartile range and n is the number of observations. Mann–Whitney rank sum tests were performed to measure statistical differences between median values of the cell types. Linear discriminant analysis was used to develop an algorithm to classify iRBCs and pWBCs based on the Hz d_eff and the RID; results were compared to the DAPI-determined cell type and the resulting area under the ROC curve was plotted.

To address the effect of the relative ratio of iRBC to pWBC on the AUC, the AUC was measured for datasets with an iRBC:pWBC relative ratio of 1:10. To achieve this, the iRBC dataset was randomly subdivided into ten equal-sized subsets. The AUC was measured for each iRBC subset compared to the full pWBC data set, giving a range of AUC depending on the data subset measured. This same process was repeated for ten datasets with an iRBC:pWBC relative ratio of 10:1.

The ability to discriminate between iRBCs and pWBCs in vivo was evaluated in a mouse model of malaria. Three cohorts of mice—negative controls (n = 3), acute infection (n = 4), and post infection (n = 3)—were included in the in vivo assessment. Female albino B6 mice were acquired from Jackson Laboratories at 7 weeks of age. Malaria infection was generated by i.p. inoculation of P. yoelii XNL (MRA-593, MR4, ATCC Manassas Virginia). Mice with acute infections, i.e. without pWBC present in the blood, were used to assess Hz signal originating from iRBCs. Mice were imaged post infection to measure Hz signal originating from pWBCs only. The surgically accessed reflected dorsal skin flap was used to mimic the tissue environment of a human mucous membrane. Hz signal was detected using a microvascular microscope (MvM), a brightfield microscope that was previously demonstrated to detect Hz in vivo and is described elsewhere [1]. For these investigations, cross polarized epi-illumination was used to optically access vessels below the tissue surface. A block diagram of the MvM and representative images collected using a non-infected mouse are shown in Fig. 2. The MvM was brought into gentle contact with the exposed vasculature of the skin flap for image collection. Linearly polarized light was delivered to the tissue by epi-illumination. Specular reflection was rejected by a polarizing beam splitter (Thorlabs, PBS201). Diffusely scattered depolarized light was imaged by a microscope objective (Newport, M-60x), relayed through a series of lenses, and focused on a CCD sensor (Point Grey, Chameleon). A liquid lens (Varioptic, Caspian) was used to manipulate the working distance of the system such that vessels at varying depths could be optically accessed without introducing motion artifact. A green-wavelength LED (λ = 532 nm) was used to locate vessels and a red-wavelength LED (λ = 655 nm) was used to detect Hz circulating within the superficial vessels. A block diagram of the MvM and representative images collected using a non-infected mouse are shown in Fig. 2. Blood samples were acquired immediately following each in vivo imaging session for thin film preparation. The presence of only iRBCs or pWBCs was confirmed by traditional blood smear microscopy. Videos collected from non-infected mice were used as negative controls.

Videos collected under red light illumination were analysed in ImageJ. A 3 × 3 neighbourhood median filter was applied to each video to reduce noise. Hz particles were identified visually as dark structures against the light background within the vessel boundary. Only Hz particles that were easily discerned in a single frame without the aid of video playback and observed in at least five consecutive frames were analysed. A 15 μm diameter region of interest (ROI) was centered on the Hz particle in a single frame. Each ROI represents a single cell. Background intensity was measured for two frames before and after the selected frame to be analysed. Similar to blood smear microscopy analysis, the intensity threshold was selected as the average minimum background pixel intensity of the ROI over time. This threshold was applied to the ROI to isolate the Hz signal, as shown in the flow chart in Fig. 3. Hz area and pixel intensity were measured for each ROI to calculate Hz d_eff and RID respectively. Additionally, the particle velocity was measured over 5 frames. Measured Hz d_eff and RID were plotted for all ROIs by infection status—negative, acute (iRBCs), or post (pWBCs) infection—as determined by blood smear analysis. Box plots characterizing the first quartile, median, and third quartile with whiskers extending to ± 1.58 (IQR/(n^1/2)) were generated for these data. Mann–Whitney rank sum tests were performed to measure whether differences in the median values of these parameters were statistically significant. ROC curves were generated using Hz d_eff and RID to discriminate between iRBCs and pWBCs and classify acute infection versus post infection. The area under the ROC curves was measured for comparison with classification performance in vitro of iRBCs and pWBCs.

Hz d_eff and RID were measured for each cell and correlated with the morphologically determined cell type from the corresponding DAPI image. Scatter plots of Hz d_eff and RID grouped by cell type are shown in the left column of Fig. 4. The right column of Fig. 4 shows box and whisker plots of these parameters for each cell type. Significant differences between median values of cell types are indicated on these plots. Median (IQR) values of Hz deff were 1.24 (0.64–1.51) for young trophozoite iRBCs, 1.48 (1.14–1.80) for late trophozoite iRBCs, and 2.94 (1.94–3.49) for pWBCs. RID median values were 7.84 (4.67–9.43) for young trophozoites, 9.27 (6.56–14.10) for mature trophozoites, and 19.62 (13.18–23.70) for pWBCs. The median RID and Hz d_eff values of iRBCs were significantly different (p < 0.0001) from pWBCs. The median values of negative control cells, uninfected RBCs and pigment-free WBCs, were significantly different (p < 0.0001) from the Hz-containing cells, iRBCs and pWBCs, for both parameters. Additionally, the RID and Hz d_eff median values were significantly greater for late trophozoites than young trophozoites (p < 0.05), suggesting the potential to estimate parasite maturity using these features.

Figure 5 shows the ROC curves for an algorithm to discriminate between iRBCs and pWBCs using Hz d_eff and RID, alone and in combination. The area under the curve (AUC) for Hz d_eff and RID were 0.89 and 0.85, respectively. The AUC for the combined parameters was 0.92. It is important to note that all levels of parasite maturity are readily observed during P. yoelii infection, however in P. falciparum infection sequestration of mature parasites decreases the number of mature iRBCs found in peripheral blood circulation. If late trophozoites are removed from the data set, the AUC increases to 0.92, 0.95, and 0.98, for each parameter respectively. Conversely, if young trophozoites are removed from the data set and only late trophozoites are compared to pWBC, the AUC decreases to 0.87, 0.80, and 0.90 for Hz d_eff, RID, and the combination of the two respectively.

Because of the wide variation in reported pWBC concentration, it is important to consider the relative ratio of iRBC to pWBC when assessing the ability to discriminate by cell type. For example, in all the cases reported in Additional file 1: Table S1, parasitaemia is at least eight times greater than pWBC concentration. The AUC was measured for ten randomly selected data sets with a relative ratio of 10:1 iRBC to pWBC in each set, yielding AUC values ranging from 0.73 to 1.00 for Hz d_eff, 0.73 to 0.92 for RID, and 0.75 to 1.00 for the combined metrics. Similarly, the AUC was measured for ten randomly selected data sets with a relative ratio of 1:10 iRBC to pWBC in each set, achieving AUC values ranging from 0.81 to 0.95, 0.64 to 0.97, and 0.80 to 0.99 for Hz d_eff, RID, and the combined classifiers, respectively.

The Hz signal features identified in blood smear microscopy were evaluated in a mouse model to classify cell type based on analysis of images of circulating cells acquired in vivo. Prior to imaging, subjects were categorized as current acute infection having iRBCs only or as post infection with pWBCs only, confirmed by blood smear microscopy. Circulating Hz was visually observed under red light illumination as small dark particles flowing through the blood vessel. The intensity and area of each particle was measured and used to calculate the Hz d_eff and RID. Randomly selected ROIs from uninfected subjects were used as negative controls. Hz d_eff and RID values measured in vivo are plotted in the left column of Fig. 6. These data show a similar trend to the blood smear data, with larger Hz d_eff and RID values for pWBCs in post infected animals as compared to iRBCs in animals with acute infections. Box and whisker plots of these parameters are shown in the middle column of Fig. 6. Median (IQR) values of Hz d_eff measurements were 2.17 (1.79–2.94) for iRBCs and 3.70 (2.31–4.68) for pWBCs. RID median values were 1.44 (1.25–2.36) for iRBCs and 4.90 (3.02–5.89) for pWBCs. Significant difference between median values of iRBCs and pWBCs were observed for both Hz d_eff and RID, p < 0.001 and p < 0.0001, respectively. The right side of Fig. 6 shows the ROC curve for an algorithm to differentiate between iRBCs and pWBCs based on these parameters. In vivo, RID performed better than Hz d_eff as a classifying feature, with an AUC of 0.91 and 0.73, respectively. Combination of the features resulted in an AUC of 0.83. As seen in the analysis of the P. yoelii-infected blood smear data, the relative ratio of iRBCs to pWBCs will affect the AUC. For a relative ratio of 1:10 iRBC to pWBC, the AUC varies from 0.61 to 0.85 for Hz d_eff, 0.79 to 1.00 for RID, and 0.84 to 0.90 for the combined classifiers in vivo. For a relative ratio of 10:1 iRBC to pWBC, the AUC varies from 0.27 to 0.99 for Hz d_eff, 0.78 to 0.99 for RID, and 0.54 to 0.91 for the combined classifiers. Variability in the Hz d_eff measurements may arise from cells coming in and out of the focal plane as they translate through the vessel. Additionally, only pWBCs occurring post-infection were investigated in vivo; the Hz d_eff may be larger in pWBCs present during an active infection when more Hz is available to be phagocytized.

The particle velocity is another feature of the Hz signal which is readily measured in vivo. This is illustrated in an overlay of successive frames from a video captured under red-light illumination in Fig. 7b. Here, a single Hz particle is tracked as it travels through the bottom vessel. In the vessel at the top of the frame, a static cell does not move over the period of 0.7 s, suggesting cytoadhesion to the endothelial wall. The same FOV imaged under green-light illumination is shown to clarify vessel location in the red illumination overlay (Fig. 7a). Average cell velocities were measured over five consecutive frames for each ROI; results are shown in the scatterplot in Fig. 7c. There is a distinct separation between the slow moving pWBCs and faster circulating iRBCs. This information could be helpful to indicate cases of cytoadhesion and sequestration in vivo and shows promise as a potential classification feature to be validated in human subjects.

Plasmodium falciparum-infected blood smears were used to evaluate the Hz signal discriminating features in a human infecting species of malaria parasites. Although these samples provided iRBCs for analysis, pWBCs were not readily detected in the smears. The lack of pWBC observation may be due to the timing of infection or very low concentrations of pWBCs. Hz is only phagocytized by neutrophils and monocytes, which account for a fraction of the total number of blood cells in a microlitre of blood; for comparison, typical concentrations of neutrophils range from 1800/µl to 7000/µl and monocytes range from 100/µl to 800/µl, whereas the concentrations of RBCs range from 4.0 to 5.5 million/µl [18], Also, the fraction of pWBCs observed in malaria infected individuals varies widely [19]. To investigate the Hz signal in pWBCs, phagocytosis of Hz isolated from P. falciparum cultured samples was induced in whole blood collected from healthy human donors. After exposure to Hz, thin smears were prepared from the blood samples and examined by conventional microscopy. DAPI signal was again used to differentiate cell type by nuclear morphology.

The Hz d_eff and RID measured in P. falciparum blood smears are plotted by cell type in the left column of Fig. 8. Only immature iRBCs were detected in the patient-derived smears, consistent with sequestration of mature iRBCs in P. falciparum pathology. Box and whisker plots of these parameters are shown in the right column of Fig. 8. The median (IQR) values of Hz d_eff were 0.52 (0.28–0.86) for iRBCs and 2.40 (1.50–3.98) for pWBCs. RID median values were 1.27 (0.72–4.30) for iRBCs and 20.60 (10.63–29.33) for pWBCs. The differences between the median values of iRBCs and pWBCs were statistically significant (p < 0.0001) for both RID and Hz d_eff. The median values of the negative control cells without Hz (uninfected RBCs and WBCs) were also significantly different from all other groups for both parameters.

These features were used to discriminate between iRBCs and pWBCs; resulting ROC curves are shown in Fig. 9. The AUC for Hz d_eff and RID were 0.94 and 0.93 respectively. When these features are combined the AUC is 0.94. These values are similar to those obtained using these features to discriminate between young trophozoite iRBCs and pWBCs from the P. yoelii blood smear data. AUCs were measured for ten randomly selected data sets with a relative ratio of 1:10 iRBCs to pWBC, yielding AUCs ranging from 0.89 to 0.99 for Hz d_eff, 0.86 to 1.00 for RID, and 0.88 to 0.99 for the combined classifiers. When the relative ratio was set to 10:1 iRBC to pWBC the AUC ranged from 0.74 to 1.00, 0.80 to 1.00, and 0.74 to 1.00 for Hz d_eff, RID, and the combined classifiers, respectively.