BACKGROUND
Malaria has plagued tropical, developing countries over the last two centuries. Light Microscopy is the gold standard for malaria parasite detection. This research work is aimed to harness the potential of virtual microscopy and computer aided diagnosis systems to minimize human error and labour towards malaria parasite detection.
OBJECTIVE
The proposed method is tested on differently stained blood smear images for malaria parasite detection.
METHODS
Digitized thin blood smears have been used to predict the presence of malaria parasite using unsupervised and rule based methods. A dataset consisting of 1410 images (667 infected, 743 normal) was developed from the MaMic database. Cochrane’s sample size estimation was used to decide sample size. To widen the applicability of the algorithm beyond the dataset under consideration, illumination correction, database specific artefact removal was performed. Thereafter unsupervised k-means (k=3) clustering was performed to segregate the foreground components, the erythrocyte, malaria infection and white blood cells from the background. Clumps are identified based on the third quartile bound of the area distribution of the foreground components. The clumps consist of both, red blood cell clumps and mixed clumps consisting of both red and white blood cells. Clumps marked out were de-clumped automatically using modified watershed algorithm. The binary de-clumped mask was used to retrieve pixel colour information from the original image. The image colour in RGB colour space was down sampled by representing the same in YCbCr colour space. Based on the values in YCbCr colour space, the image was recoloured and pixel position matching was performed to detect malaria parasite.
RESULTS
As compared to Zack’s thresholding (63.75%), 3-means clustering (98.96%) had a higher accuracy at foreground particle identification. The third quartile mark was selected for clump/s identification while Tukey’s upper hinge showed higher strength towards white blood cell particle identification. The accuracy for malaria parasite detection by the proposed system was recorded as 98.11% (Sensitivity-0.9645, Specificity-1, AUC-0.9583)
CONCLUSIONS
The proposed work is particularly innovative as it uses two basic features, colour and area, to identify malaria parasite in thin blood smear image. The paper documents an automated robust algorithm to assist pathologists at Parasitaemia estimation as per World Health Organization standard.
A Novel Implementation of Deep-Learning Approach on Malaria Parasite Detection from Thin Blood Cell Images
