Fig 1.
Prototype point-of-care digital slide scanner.
(a) The prototype digital slide scanner (1) with slide adjustment motor unit (2) and laptop computer used to control device (3). (b) USAF standardized resolution test chart, digitized with the microscope prototype, showing the smallest set of resolvable bars (corresponding to a spatial resolution of 0.9 μm).
Fig 2.
Microscopic field of view (FOV), showing the corresponding (a) brightfield, (b) fluorescence and (c) merged thin blood smear digital images. Red bounding boxes showing enlarged areas with (1) infected red blood cells (RBCs), (2) normal RBCs, (3) leukocytes and (4) fluorescent debris.
Fig 3.
Training of deep-learning system 1 and analysis of blood smear.
Workflow for training of the first deep-learning system (DLS 1) and subsequent analysis of samples using the trained model. (a): Training of the DLS was performed on merged images (brightfield and fluorescent images combined) where regions with visible red blood cells (RBCs) and trophozoites (parasites) were manually annotated and used to train the DLS to detect RBCs and classify them as infected vs. non-infected. (b): Analysis of samples in the validation series was performed on merged images in two steps: 1) Segmentation of all visible RBCs and 2) quantification of infected and non-infected RBCs to determine overall level of parasitaemia.
Fig 4.
Training of deep-learning system 2 and analysis of blood smear.
Workflow of training and analysis with the second deep-learning system (DLS 2) using the GoogLeNet model. Panels showing brightfield images with segmented red blood cells (RBCs), the corresponding fluorescence image with detected parasite candidates, classification of parasites and exported analysis results.
Table 1.
Results from analysis of blood smears with the different methods.
Fig 5.
Detected level of parasitemia by the different methods.
Box plots illustrating the detected levels of P. falciparum parasitemia (percentage of infected red blood cells; RBCs) in malaria-positive blood smears collected prior to initiation of treatment (Day 0) by the deep learning systems (DLS) and by visual analysis of the digital, DAPI-stained slides and by conventional light microscopy of Giemsa-stained thick smears.
Fig 6.
Correlation between detected levels of parasitemia by deep learning-based and visual sample analysis.
Correlation between levels of P. falciparum parasitemia, as detected with the deep learning systems (DLSs), and by visual analysis of the digital samples, collected on (a) Day 0 and (b) Day 3).
Fig 7.
Detected levels of parasitemia on Day 0 and Day 3, following initiation of treatment.
Box plots illustrating the detected levels of P. falciparum parasitemia (percentage of infected red blood cells; RBCs) in blood smears collected at baseline, prior to initiation of treatment (Day 0) and three days following initiation of treatment (Day 3). Results shown as detected by analysis with the deep learning systems (DLS) and by visual analysis of the digital slides.