Automated Recognition of Plasmodium falciparum Parasites from Portable Blood Levitation Imaging

Abstract

In many malaria-endemic regions, current detection tools are inadequate in diagnostic accuracy and accessibility. To meet the need for direct, phenotypic, and automated malaria parasite detection in field settings, a portable platform to process, image, and analyze whole blood to detect Plasmodium falciparum parasites, is developed. The liberated parasites from lysed red blood cells suspended in a magnetic field are accurately detected using this cellphone-interfaced, battery-operated imaging platform. A validation study is conducted at Ugandan clinics, processing 45 malaria-negative and 36 malaria-positive clinical samples without external infrastructure. Texture and morphology features are extracted from the sample images, and a random forest classifier is trained to assess infection status, achieving 100% sensitivity and 91% specificity against gold-standard measurements (microscopy and polymerase chain reaction), and limit of detection of 31 parasites per µL. This rapid and user-friendly platform enables portable parasite detection and can support malaria diagnostics, surveillance, and research in resource-constrained environments.

Keywords: computer vision; malaria; portable imaging; resource-limited settings.

Figure 1

Rapid sample suspension and portable imaging approach for parasite detection. A) Schematic depicting the life cycle of the parasite Plasmodium falciparum, describing the stages of the asexual blood‐stage forms, where the ring‐stage parasite is the focus of clinical detection. B) Schematic describing mechanism of sample preparation in levitation chamber. A 10 µL drop of whole blood is mixed with a temperature‐stable solution (saponin and paramagnetic ions in phosphate buffered saline (PBS)) and loaded into a square glass capillary (1 µm tall). The contents are suspended for 3 min, in which the red blood cells (RBCs) lyse to release free hemoglobin and, if infected, intracellular parasites, and all objects in suspension (including unlysed white blood cells (WBCs)) come to equilibrium positions. C) Custom prototype for portable imaging setup of levitated blood samples. Sample capillary is inserted as indicated, backlit with a white light emmitting device (LED), for brightfield imaging, and the magnified and corrected image is viewed on the smartphone screen as captured on its camera. This intuitive display and touchscreen can be used to capture images and focus can be adjusted. The modular design allows for changing magnification and adding fluorescence imaging. D) Example images of whole blood samples from human donors as captured on the portable imaging platform, marked with regions where WBC are present, and where parasites are present: (i) an example of a malaria‐negative blood sample, and (ii) an example of a malaria‐positive blood sample. Scale bar = 100 µm. E) Image analysis and classification workflow summary. Abbreviations:  iRBCs = infected RBCs, uRBCs = uninfected RBCs.

Figure 2

Investigating erythrocyte lysis to aid direct visual parasite detection. A) Different approaches to lysis of RBCs for detection of parasites, with the symbols above indicating suitability for resource‐limited settings, as discussed in the text: (i) Streptolysin O, (ii) saponin, (iii) Tween 20. B) Dilution of culture‐grown P. falciparum under saponin lysis (n = 3 blank samples, n = 10 parasite‐containing samples) in levitation suspensions at 50 mm of gadolinium, to determine the limit of detection (31 parasites per µL). Actual parasitemia is calculated from the known parasitemia of the culture and the dilution factor, while detected parasitemia is calculated from the concentration of parasites detected visually in each volumetric image field. Detected parasitemia correlated to actual parasitemia with R ² = 0.882. C) (i) Example microscope image of whole blood spiked with cultured parasites, subject to the saponin lysis protocol. Acridine orange stains nucleic acids green, thus highlighting the two bands of parasites (released from RBCs) and WBCs (unlysed) respectively. (ii) The box‐and‐whisker plot shows the mean height of bands in samples of lysed parasites from culture, that is, parasite bands (n = 5) and lysed uninfected whole blood, that is, leukocyte bands (n = 5). Both groups were significantly different under a t‐test with a p‐value of 0.0089 (two asterisks). Each element in the box plot represents the following: center line represents the median; box limits represent the upper and lower quartiles; whiskers represent the range of the data. D) We selected the saponin lysis protocol to use with malaria‐infected and uninfected human whole blood samples. Close‐up images of saponin‐lysed whole blood samples from human donors, as captured on the portable imaging platform, to show cell and lysate structures in more detail: (left) an example of a malaria‐negative blood sample, and (middle) an example of a malaria‐positive blood sample. The orange–red background color of the solution is attributed to free hemoglobin from lysed RBCs, the light mottled pattern in the background is attributed to ghosted RBC membranes, and the small dark circular structures in the foreground are attributed to the intracellular ring‐stage parasites. Scale bar = 10 µm. (Far right): higher‐zoom images showing examples of ring‐stage morphology. Scale bar = 5 µm. Abbreviations: RBCs = red blood cells, WBC = white blood cells. Symbols in (A), from left to right: sensitive, rapid, temperature‐stable, low‐cost, electricity‐independent, user‐friendly.

Figure 3

Automated image processing to extract visual features associated with parasites. A) A Laplacian‐of‐Gaussian (LoG) blob detection algorithm developed to directly and specifically detect the ring‐stage parasites, and differentiate them from other structures (the background, which can contain ghosted RBC membranes, as well as intact WBCs which are larger and of a different shape). Scale bar = 10 µm. B) The results of the LoG algorithm to quantify parasite structures in the patient sample images acquired on the portable platform. The quantification is the number of blobs detected per sample image. Performing a z‐test on both groups resulted in a test statistic of −2.2 and a p‐value of 0.025 (marked with one asterisk). Using a threshold of 10, sample images can be classified into negative (below 10) or positive (greater than or equal to 10). C) A Bland–Altman analysis was conducted to compare the parasitemia quantification accuracy of this algorithm to the gold standard method of expert‐examined Giemsa‐stained smear microscopy. D) Image analysis of texture features was performed on gray‐level co‐occurrence matrixes (GLCM) of the image pixels, using Haralick's method to extract 13 specific features calculated from the relationships between pixel structures in the GLCM. These features were calculated, per image, for four directions each, and the difference therein between malaria‐negative and malaria‐positive samples was explored. The directional means of the five features that contained the most predictive differences between the two groups are presented here: (i) angular second moment (test statistic: 2.7, p‐value: 0.0076), (ii) contrast (test statistic: −3.1, p‐value: 0.0020), (iii) inverse difference moment (test statistic: 4.4, p‐value: 0.000013), (iv) difference variance (test statistic: 4.5, p‐value: 0.0000057), and (v) first informational measure of correlation (test statistic: −4.9, p‐value: 0.00000076). Notes on the box‐and‐whisker plots: Each element represents the following: center line represents the median; box limits represent the upper and lower quartiles; whiskers represent the range of the data, circles represent outliers in the data. Abbreviations: RBCs = red blood cells, WBC = white blood cells.

Figure 4

Building a machine learning classifier of infection status using image features. A) A visualization of the random forest classifier applied to our available dataset, which consisted of 53 features (13 Haralick features, each in four directions, and the number of parasite‐like objects detected in each image by the Laplacian‐of‐Gaussian (LoG) blob detection algorithm) for a total of 81 samples, which were randomly split into a training set (75%, or 60 samples) and a test set (25%, or 21 samples). This classifier used ten estimators, all of which can be seen in the supplement. B) Of the 53 features used, the 10 most important are shown here, along with their relative contributions to the performance of the classifier. C) A confusion matrix quantifies the performance of this classifier on the training set, with 100% sensitivity, 96% specificity, and 98% accuracy overall. D) A confusion matrix quantifies the performance of this classifier on the test set, with 100% sensitivity, 91% specificity, and 95% accuracy overall. E) A receiver operating characteristic (ROC) curve is drawn for the classifier, on both training and test sets, with both showing high area under the curve (AUCs) and consistent performance between the training and test set. The training set AUC was 1.00 and the test set AUC was 0.991. Abbreviations: RBCs = red blood cells, WBC = white blood cells.