Malaria diagnosis based on a machine learning system
Chidinma Maduako · eScholarship@McGill (McGill) · 2020
SUMMARYIntroduction: The latest World Malaria Report released in November 2017 estimated that 219 million cases of malaria occurred and deaths due to malaria reached 435,000 in 2017(1). The WHO considers microscopy to be the gold standard for clinical diagnosis of malaria due to its ready availability. However, microscopy has many shortcomings, including inter-user variability and inconsistency, due to the fact that many microscopy technicians do not assess the standard number of high-power fields, are not adequately trained on recognizing all forms of malaria and the high disparity associated with the quality of manual Giemsa slide production (4) To remedy the mis-use of empiric (symptom-guided) treatment, malaria testing is required by many governmental health organizations before commencing antimalarial drug therapy, thereby resulting in increased demand for up to 500 million malaria tests in 2012 (10). Understanding the diagnostic expertise necessary and representing it by specifically tailored image processing, analysis and pattern recognition algorithms can help in designing an automated diagnosis system. Although it is not yet a widespread research topic, automated diagnosis of malaria directly addresses several current gaps (11). My research aims to develop a machine learning system that can identify the stage and number of Plasmodium falciparum in cultured erythrocytes based on their morphology using thin film slides, the second objective is to develop a machine learning system that can identify the number of ring-stage parasites in samples of cultured erythrocytes diluted with fresh whole blood.Methods: Giemsa stained thin blood smears were made from synchronized cultures of the 3D7 strain of plasmodium falciparum stored in an incubator with shaking at 37oC, 5% CO2, 3 % O2, 92 % N2. Thin blood smears were viewed with an EVOS microscope and digital images were acquired, saved as tiff format and stored in a memory stick. The images were transferred as files to a computer, then the images further pre-processed, segmented, and the parasites and the stages of the life cycle detected. The algorithm formulated with the MATLAB programme was trained using 109 images. For the first objective, 397 images were used and for the second objective 163 images were used.The Otsu algorithm was used for this study, gray level images were reduced to binary images. The algorithm assumes that the images contain foreground and background pixels. Results: This study showed a relatively strong, positive linear association/ correlation between automated count and manual counts. The correlation between the manual count and the automated count was 0.85. The Pearson correlation between the automated and manual count was 0.7. The diagnostic tool showed a sensitivity of 94.6% for rings, 96.5% for trophozoites and 98.2% for schizont. Moreso, it showed a specificity of 96.5% for rings, 88.9% for trophozoites and 81.8% for schizonts. The R and G channels of the RGB color scheme had clear features which were used to identify objects containing chromatin in Giemsa-stained blood films. The input images transformed to grayscale highlighted parasites containing chromatin.Conclusion: This study developed an automated system that could enhance the diagnosis and therefore treatment of malaria. The automated method detected more trophozoites and schizonts than the ring stage parasites as seen with a correlation value of 0.83, 0.86 and 0.94 for the ring, trophozoite and ring stages respectively