You are here
Cascading YOLO: Automated Malaria Parasite Detection for Plasmodium Vivax in Thin Blood Smears.
Purpose: Malaria continues to be a major burden on global health, causing approximately half a million
fatalities each year . However, malaria diagnosis from patient blood smears requires a significant amount
of time and expertise. Malaria is caused by Plasmodium parasites including five species, of which
Plasmodium falciparum and Plasmodium vivax pose the greatest health threat . The goal of this work is
to develop an automated malaria parasite detection system for Plasmodium vivax. The automated and
accurate detection of Plasmodium vivax is made difficult due to the lower parasitemia levels typically
observed in these infections, as compared to Plasmodium falciparum, which tends to require microscopists
to image regions of thin blood smears that are dense with cells and challenging to segment.
Contributions: Our main contribution is two-fold: First, in the literature, only a few papers have been
published on the automated identification of Plasmodium vivax [3-5], and we are the first to perform
evaluation on patient level. Second, we propose a cascade of YOLO model and AlexNet classifier for hardnegative
mining to reduce the false positive errors of the parasite detection, which improves the mean
average precision by about 8%.
Methods: We develop a rapid and robust diagnosis system for the automated detection of Plasmodium
vivax parasites using a cascaded YOLO model. This system consists of a YOLOv2 model  and a classifier
for hard-negative mining. We first split images of 4032×3024×3 pixels into regions of 672×504×3 pixels,
and train a YOLOv2 model with the regions and corresponding manual ground-truth annotations. To reduce
false positive errors, we cascade a transferred AlexNet classifier trained with annotated parasites and false
positives (hard-negatives) generated by the YOLOv2 model. For testing, we also split blood smear images
into regions of 672×504×3 pixels, which are then screened for parasites using our cascaded YOLO model.
The detected parasite coordinates are then re-projected to the original image space for visualization and