Figure 1.
A flow chart of the proposed P. falciparum detection method resulting in a panel showing the most probable detections in one thin blood smear as well as the parasitemia count for the same sample.
Figure 2.
The candidate region segmentation phases: A) a typical bimodal histogram and definition of thresholds TS and TB, B) an example tile segmented into C) foreground and background, D) strongly stained regions based on the histogram and E) the remaining candidate region after filtering out the smallest and the largest objects.
Figure 3.
The sliding window support vector machine classification illustrated by A) an example tile showing all the windows extracted from the tile based on the candidate region segmentation, B) the resulting heat map after classification, in which the decision scores for the windows are visualized as small squares colored according to a color map and C) a bar showing the heat map for the classifier score values.
Figure 4.
The reader assessment mode displayed on a tablet computer.
A) A portrait view showing the highest ranked P. falciparum detections, B) the highest ranked detections and their corresponding heat maps viewed in a landscape mode, C) an example of a strong detection highly suspicious for P. falciparum and D) a lower ranked detection not suspicious for P. falciparum, E) a panel of the highest ranked areas displayed on a tablet computer.
Figure 5.
Parasitemia is estimated by first segmenting a set of erythrocytes and then scoring them based on the heat map e.g. the sliding window classification results.
A) An example tile, B) segmented erythrocytes, C) the corresponding heat map, D) a bar showing the heat map for the classifier score values and E) the resulting erythrocyte classification to either P. falciparum infected (red) and uninfected erythrocytes (blue).
Figure 6.
Ground truth annotations and outcome of the computer vision-assisted decision support method in each digitized thin blood film in the test series (19 malaria infected samples and 12 uninfected controls).
Thumbnail pictures show for each patient ten of the 128 sample areas (i.e. detections) with highest probability of malaria infection detected by the image analysis algorithm and presented to the expert in the panel view described in Figure 4. Parasitemia was only calculated for cases considered as malaria positive based on visual inspection of the highest scoring detections. Note that only part of the erythrocytes were successfully segmented by the algorithm and therefore also the ground truth annotations in the segmented cells is lower than the total number of annotated parasites in a sample.
Figure 7.
The agreement between the estimated parasitemia and ground truth is shown with A) a logarithmic agreement plot and B) a Bland-Altman plot.