The Malaria Segmentation: A Dataset for Segmentation of Malaria Infected Cells consists of Giemsa stained images derived from blood samples of 16 patients infected with Plasmodium falciparum. Skilled experts meticulously annotated the different types of parasites within each image. The authors adopted a two-step methodology, which began with image segmentation and was followed by life stage classification. During the segmentation phase, each pixel was categorized as either a parasite or a non-parasite pixel using a random forest classifier. The performance of this segmentation process was assessed using metrics like classification accuracy, Dice coefficient, and free-response receiver operating characteristic (FROC) analysis. For the life stage classification, the segmented objects were divided into 8 classes, covering a range of parasite life stages, as well as other elements such as white blood cells or debris.

The dataset was derived from patient samples in Gambia. Plasmodium falciparum parasites were cultured ex-vivo for 24 to 48 hours. Giemsa-stained images revealed the parasites as dark purple entities, contrasting with the light pink appearance of red blood cells. Each image consisted of 4 by 4 sub-images, which were separated for individual analysis. The dataset comprised a total of 837 images obtained from 16 patients, with one patient omitted due to flawed staining.

To establish ground truth image segmentation, the authors manually selected and cropped images around dark-stained blobs containing parasites, white blood cells, or debris. Initial segmentation of parasites was performed using Otsu thresholding on the green channel, followed by manual adjustments and fine-tuning. Morphological operations such as imfill() and imdilate() in Matlab were employed for further refinement. The segmented cropped images were combined into a ground truth binary image.

The parasite life stages were described as follows. R = light ring, LR/ET = fat rings with dark cytoplasm, very small trophozoites with pigment; rif stage, MT = pigment, vacuole, mid-size, LT = two nuclei, very large, Esch = more than two nuclei, dark staining, very large, Lsch = still has RBC, but clear merozoites, Seg = no RBC (or very faint outline); bunch of grapes. Debris is a left-over category for those blobs that could not be clearly classified as parasites or WBCs. There were a total of 2911 parasites in different life stages, with relatively limited class imbalance. While the focus was not on WBC or debris classification, their presence was taken into account within the methodology.