Abstract
BACKGROUND: Detection of malaria parasite from blood smears remains the gold standard for confirmation of diagnosis. Screening blood smears for malaria parasite has a sensitivity of 75 %, and requires intensive training of the laboratory technician. In the present study, we have attempted to develop an artificial intelligence to automate the process of malaria parasite detection. METHODS: We acquired 352 images of Leishman-Giemsa-stained peripheral blood smears, containing either normal red blood cells (RBCs) or parasitised RBCs. With a trial and error approach, we developed five deep learning models: (A) Naive deep convolutional neural network (DCNN) for trophozoites, (B) Modified Inception V3 pretrained neural network (C) Combination of model A and B, (D) Segmentation of cells from the images through Watershed Transform and naive tri-class DCNN (normal RBCs, parasitised RBCs, WBC/platelets), and (E) A naive DCNN model to detect ring forms. The images were randomly split into training and test sets and training was imparted on all the models. After completion of training, performance of each model was assessed on the test set. RESULTS: Overall, the best combination of sensitivity and specificity was seen in model D (85 % and 94 %, respectively) in detecting parasites; in addition to trophozoites, model D could also detect ring forms. The performance of model A, B & C suffered from lack of either sensitivity or specificity. CONCLUSION: The present study represents the first step towards development of a complete module for screening malaria parasites from automated microphotography/whole slide images.